Spontaneous vs Posed: Analysis of Emotion Intensities using Wavelet Coefficient and Support Vector Machine

Spontaneous vs Posed: Analysis of Emotion Intensities using Wavelet Coefficient and Support Vector Machine | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Spontaneous vs Posed: Analysis of Emotion Intensities using Wavelet Coefficient and Support Vector Machine Asraful Syifaa' Ahmad, Rohayanti Hassan, Noor Hidayah Zakaria, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5379043/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Emotion detection is a critical aspect of human-computer interaction and various other applications, including real-time video surveillance, psychological studies and social media emotion analysis. Understanding the complexity and nuance of human emotions requires an in-depth look at emotion intensity. Furthermore, occlusion is a common problem in facial emotion detection, leading to low detection accuracy. This paper presents a unique method for utilising Support Vector Machine (SVM) classification with wavelet coefficient analysis to analyse emotion intensities. To assess its effectiveness, we tested our approach on several benchmark datasets, such as the Japanese Female Facial Expression (JAFFE) dataset and the Facial Expression Recognition 2013 (FER2013) dataset. Our method demonstrates scalability and adaptability for application in large datasets, addressing the needs of big data environments. Additionally, we tested our method on occluded sections of the FER2013 dataset to simulate real-world conditions where facial emotions might be partially obscured, such as in video surveillance settings. Results indicate that wavelet coefficient analysis successfully captures nuanced differences in emotional intensity, while SVM performs robustly in classifying emotions, even in challenging, partially occluded conditions. Furthermore, the system’s structure allows for effective implementation in real-time applications and adaptability to the vast data streams typical of social media and large-scale psychological datasets. Our findings suggest that the proposed framework improves detection processes and offers a scalable, adaptable solution for large-scale emotion analysis across varied applications and significant data contexts. Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Emotion detection is a process of recognising and categorising human emotion from any source, including verbal and non-verbal communication. Facial expression and body language, including body posture and gesture, are non-verbal forms of communication (1,2). Emotion represented by facial can be divided into six discrete categories (3,4). These six basic emotions are Happiness, sad, anger, fear, surprise, and disgust. Thus, previous researchers have investigated facial features vigorously for emotion analysis (4–6). Emotion detection has various applications, including mental health detection for conditions like depression, posttraumatic stress disorder (PTSD), and autism. It also benefits online education, allowing teachers to monitor and understand their students' emotions. In addition, the entertainment industry can utilise emotion detection to enhance user experiences, and e-commerce platforms can leverage it to gauge customer emotions. For instance, e-commerce platforms often provide rating and ranking systems where customers can express their emotions and satisfaction levels regarding the services they receive from providers. This information encoded in customer reviews can provide valuable insights into their emotions related to the purchase experience (7–9). Moreover, emotion detection has broad applications in diverse fields, such as real-time video surveillance systems, large-scale psychology studies, and social media sentiment analysis, where accurate identification of emotional states is crucial for interpreting human behaviour in various contexts. Facial emotion detection and facial expression recognition are interconnected areas that involve the analysis of facial expressions to understand a person's emotions. While these terms are occasionally used interchangeably, facial emotion detection primarily focuses on recognising general emotional states, while facial expression recognition aims to identify specific expressions. Both fields utilise sophisticated algorithms and machine learning techniques to analyse facial features and find applications in mental health assessment and human-robot interaction. However, despite technological advancements, facial emotion detection still faces significant challenges, such as occlusions, subtle expression variations, illumination differences, and low-resolution images, all affecting detection accuracy. These issues become even more critical when scaling detection methods to large datasets. Given these needs, our study proposes an improved facial emotion detection framework that leverages wavelet coefficient analysis and Support Vector Machine (SVM) classification to select the most significant features for each emotion, thereby enhancing detection accuracy. This approach addresses occlusion and other common obstacles and enhances system adaptability, making it suitable for high-performance applications in big data environments. To achieve this aim, this paper has conducted the following investigations: 1. Histogram Intensities Comparison Across Seven Emotions in the JAFFE Database. 2. Emotion Classification Using Support Vector Machine on the JAFFE Dataset 3. Improvement in Accuracy by Having Wavelet Coefficient with Support Vector Machine 4. Cross-Database Evaluation Performance Using JAFFE Dataset and Occluded FER2013 Dataset The remainder of this paper is organised as follows: Section 2 provides the related works according to emotion detection, and Section 3 provides an overview of the proposed method, techniques, and dataset for the whole framework. The results are presented and discussed in Section 4, while Section 5 concludes the paper. Related Works This section consists of an overview of public facial emotion datasets and recent feature extraction and classification techniques. The table provides an overview of various emotion detection datasets that have been publicly used. Each of the datasets is comprised of basic emotions captured from either posed participants, real-world situations, or video. Table 1. Dataset of the benchmark dataset Dataset Emotion Description Setting Image MMI (10) Six basic emotions & neutral 2900 videos (the neutral, onset, apex and offset) Posed BU-3DFE (11) Six basic emotions & neutral 2500 3D facial images captured on two views -45°, +45° Posed MultiPie (12) Anger, disgust, neutral, happy, squint, scream, surprise 750,000 images (15 view and 19 illumination conditions) Posed CK+ (13) Six basic emotions, contempt & neutral 593 videos (posed and non-posed expressions) Both RaFD (14) Six basic emotions, contempt & neutral 8040 images (different face poses, ages, genders & sexes) Posed GEMEP FERA (15) Anger, fear, sad, relief, happy 289 images sequences Posed SFEW (16) Six basic emotions & neutral 700 images (different ages, occlusion, illumination and head pose) Spontaneous Oulu-CASIA (17) Six basic emotions 2880 videos (three different illumination conditions) Posed FER2013 (18) Six basic emotions & neutral 35,887 grayscale images (from Google Image search) Spontaneous CASME II (19) Happy, disgust, surprise, regression & others 247 micro-expressions sequences Posed RAFD-DB (20) Six basic emotions & neutral 30,000 images from the real world Spontaneous AffectNet (21) Six basic emotions & neutral More than 440,000 images collected from the internet Posed JAFFE (22) Six basic emotions & neutral 213 grayscale images posed by 10 Japanese females Posed The recognition performance is affected by the number of features extracted as input for the classifiers (23). Thus, a suitable extraction method is needed. Feature extraction is a process in recognition that extracts important and meaningful features that could represent facial images. Later, these features will be input to the classification step (24). Facial expressions hold abundant information regarding an individual's internal thoughts and emotions, making them crucial for human communication. In real-world scenarios, the images captured contain odd expressions, illuminations, face occlusion, less resolution, different distances from the camera, and different angles of face orientation (1,25,26). Image histograms are frequently used in various image processing applications, including image enhancement, segmentation, and analysis (27–29). Histograms are graphic representations that illustrate the frequency of pixel occurrences for each intensity value in a picture, thereby displaying the distribution of pixel intensities. An 8-bit grayscale image has 256 distinct intensities, ranging from 0 to 255. These values facilitate an in-depth understanding of the image's features, wherein several peaks may suggest the presence of different objects or regions in the image or offer insights into the general illumination, differences, and range of the image. The graph displays intensity values on the x-axis and the frequency or number of pixels at each intensity value on the y-axis (30). Additionally, it can perform advanced analysis such as histogram equalisation, histogram specification, and local enhancement. The table presents significant progress in emotion detection technology, featuring several approaches, including fuzzy inference systems, advanced neural network designs, and conventional supervised learning techniques. Over time, the accurateness of these methods has significantly enhanced, with modern techniques such as Transfer Learning + pre-trained DCNN and different Improved CNN models acquiring nearly perfect accuracy. The Fuzzy Emotion Inference System, created by Liliana et al. in 2019, employs a Fuzzy Inference Engine specifically built to manage uncertainty and imprecision in the recognition of emotions. This approach obtains an accuracy rate of 82%, suggesting a moderate degree of accuracy. The Attention-Based DenseNet (MABD) technique utilises the DenseNet architecture along with attention processes to enhance the recognition of essential features for emotion detection. Initially, applications of Convolutional Neural Networks (CNNs) for emotion detection, as demonstrated by Wen et al. in 2017, exhibited a comparatively low accuracy of 45.07%. This result indicates the early stage of CNN utilisation in this field during that period. On the other hand, a more up-to-date CNN implementation conducted by Shen and Xu in 2023 exhibits an essential enhancement in accuracy, reaching 93.33%. This highlights the progress made in CNN structures and training methods. Appasaheb Borgalli and Surve developed a specialised CNN architecture in 2022 that delivers a high accuracy rate of 91.58% for emotion recognition, geared to meet specific criteria. The Deep Neural Networks (DNNs) employed by An et al. in 2023 exhibit exceptional accuracy, reaching 96.97%. This result highlights their efficacy in collecting complex patterns within emotional data. The CNN approach proposed by Wen et al. in 2017, which incorporates enhancements, demonstrates minor improvements, achieving an accuracy rate of 50.7%. Fuzzy Classification methods, exemplified by the approach developed by Nicolai and Choi in 2016, integrate fuzzy logic with classification techniques to achieve an excellent accuracy level of 78.33%. In 2019, Kim et al. developed a Hierarchical Deep Neural Network Structure that effectively improves emotion recognition by utilising a hierarchical structure. This model achieved an impressive accuracy of 91.27%. Enhanced CNN methodologies demonstrate exceptional performance. Zaman et al. achieved an accuracy of 98.01% in 2023, whereas Ramis et al. reached 98.29% in 2022. Similarly, in 2019, Wang et al. achieved an accuracy of 95.95%. In 2023, Benisha and Mirnalinee achieved an accuracy of 96.91%, while Podder et al. achieved 96.83%. These results demonstrate the exceptional performance of advanced CNN algorithms. In 2023, Chandra et al. developed a modified version of Attention-Based DenseNet (MABD), which demonstrated a lower accuracy of 42.99%. This indicates the presence of potential difficulties or areas that require additional improvement. The PSA-YOLO model, created by Zhang and Ma in 2023, employs the YOLO architecture with specific alterations to identify emotions. It has achieved an accuracy rate of 83.84%. The Stacked Sparse Auto-Encoder developed by Ahmad et al. in 2023 demonstrates a 92.5% success rate, indicating its efficacy in encoding and classifying emotions. The utilisation of pre-trained deep convolutional neural networks (DCNN) in transfer learning, as demonstrated by Akhand et al. in 2021, yields the maximum accuracy of 99.52% in the provided table, effectively showcasing the efficacy of pre-existing models. Supervised learning approaches exhibit exceptional performance. In 2010, Chatterjee and Shi developed the Adaptive Neuro Fuzzy Inference System, which effectively combines neural networks and fuzzy logic to achieve a remarkable accuracy rate of 95%. The DLSANet, developed by Guo et al. in 2023, demonstrates an impressive accuracy rate of 93.81%. The HOG+ Random Forest algorithm, developed by Saeed in 2023, attains an accuracy of 92.97%. K-Nearest Neighbours (KNN) approaches exhibit different degrees of accuracy. Shelke et al. achieved an accuracy of 77.29% in 2021, whereas Cornejo et al. achieved a far higher accuracy of 98.21% in 2015. In a study conducted by Harakannanavar et al. in 2023, they were able to attain an accuracy rate of 96.51%. In 2023, Subudhiray et al. demonstrated that combining KNN with Gabor filters, HOG features, and Local Binary Patterns (LBP) resulted in accuracies of 94.24%, 84.53%, and 90.65%, respectively. The multi-SVM algorithm, developed by Zhu et al. in 2015, exhibits a lesser accuracy rate of 61.97%. In contrast, the Softmax algorithm, proposed by Liu et al. in 2017, obtains a higher accuracy of 90%. Support Vector Machine (SVM) methods exhibit a variety of accuracy levels. In 2021, Jeen Retna Kumar et al. achieved an accuracy of 97.3% by integrating Support Vector Machines (SVM) with cross-dataset analysis and feature improvement. In 2013, Zhou et al. attained an accuracy of 92.14%, while in 2017, Liu et al. achieved an accuracy of 89.6%. Lu et al. (2019) had a success rate of 98%, Cornejo et al. (2015) achieved a success rate of 82.86%, and Harakannanavar et al. (2023) achieved a success rate of 98.26%. 2019 Eng et al. demonstrated a reduced accuracy rate of 76.19%. The study that integrates Support Vector Machines (SVM) with Wavelet Coefficients (WC) produces a high accuracy rate of 94%. In contrast, the SVM Stacking method proposed by Delazeri et al. in 2022 only achieves an accuracy of 64.59%. Frank's enhanced Support Vector Machine (SVM) technique shows a remarkable accuracy of 94.13%. The unsupervised learning methods developed by Tuncer et al. in 2023, such as PCA + LDA, show a remarkable accuracy of 97.09%. These techniques effectively combine dimensionality reduction and feature extraction. The table provides an overview of the notable progress made in emotion detection technology, demonstrating various approaches such as fuzzy inference systems, advanced neural network layouts, and conventional supervised learning methods. Over time, the precision of these methods has significantly enhanced, with modern techniques such as Transfer Learning + pre-trained DCNN and different Improved CNN models obtaining nearly perfect accuracy. This history demonstrates algorithms' continuous advancement and growing capacity to accurately represent and categorise human emotions. Table 2. Emotion detection by previous studies Method Year Accuracy Remarks Categories Fuzzy emotion inference system (31) 2019 82 6 emotions (Happy, anger, sad, disgust, fear and surprise) Fuzzy Inference Engine CNN (32) 2017 45.07 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network CNN (33) 2023 93.33 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network Custom CNN Architecture (34) 2022 91.58 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network DNN (35) 2023 96.97 4 emotions (angry, happy, neutral, fear) Neural network ECNN (32) 2017 50.7 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network Fuzzy Classification (36) 2016 78.33 6 emotions (Happy, anger, sad, disgust, fear, and surprise) Neural network Hierarchical Deep Neural Network Structure (37) 2019 91.27 6 emotions (happy, anger, sad, disgust, fear, and surprise) Neural network Improved CNN (38) 2023 98.01 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network Improved CNN (39) 2022 98.29 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network Improved CNN (40) 2019 95.95 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network Improved CNN (41) 2023 96.91 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network Improved CNN (42) 2023 96.83 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network Modified Attention-Based DenseNet (MABD) (43) 2023 42.99 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network PSA-YOLO (44) 2023 83.84 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network Stacked Sparse Auto-Encoder (45) 2023 92.5 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network Transfer learning + pre-trained DCNN (46) 2021 99.52 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Neural network Adaptive Neuro Fuzzy Inference System (47) 2010 95 5 emotions (happy, anger, sad, disgust and surprise) Supervised learning DLSANet (48) 2023 93.81 7 emotions (Happy, anger, sad, disgust, fear, surprise and neutral) Supervised learning HOG+ Random Forest (49) 2023 92.97 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Supervised learning KNN (50) 2021 77.29 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Supervised learning KNN (51) 2015 98.21 7 emotions (Happy, anger, sad, disgust, fear, surprise and neutral) Supervised learning KNN (52) 2023 96.51 6 emotions (Happy, anger, sad, disgust, fear, and surprise) Supervised learning KNN+Gabor (53) 2023 94.24 6 emotions (happy, anger, sad, disgust, fear, and surprise) Supervised learning KNN+HOG (53) 2023 84.53 6 emotions (happy, anger, sad, disgust, fear, and surprise) Supervised learning KNN+LBP (53) 2023 90.65 6 emotions (happy, anger, sad, disgust, fear, and surprise) Supervised learning Multi-SVM (54) 2015 61.97 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Supervised learning Softmax (55) 2017 90 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Supervised learning SVM (56) 2021 97.3 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Supervised learning SVM (57) 2013 92.14 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Supervised learning SVM (55) 2017 89.6 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Supervised learning SVM (50) 2021 83 7 emotions (happy, anger, sad, disgust, fear, surprise and neutral) Supervised learning SVM (58) 2019 98 6 emotions (Happy, anger, sad, disgust, fear, and surprise) Supervised learning SVM (51) 2015 82.86 7 emotions (Happy, anger, sad, disgust, fear, surprise and neutral) Supervised learning SVM (52) 2023 98.26 6 emotions (Happy, anger, sad, disgust, fear, and surprise) Supervised learning SVM (59) 2019 76.19 7 emotions (Happy, anger, sad, disgust, fear, surprise and neutral) Supervised learning SVM Stacking (60) 2022 64.59 7 emotions (joy, anger, sad, disgust, fear, surprise and contempt) Supervised learning PCA + LDA (61) 2023 97.09 7 emotions (Happy, anger, sad, disgust, fear, surprise and neutral) Unsupervised Learning Researchers must address several current issues to improve the accuracy and reliability of facial emotion detection. The first is dependent setting. The low accuracy of emotion detection using facial expressions is caused by several issues, such as the dependent setting of expression, the distance of the face from the camera, obstacles in front of the face that cause face occlusion, and the low resolution of an image. Facial occlusion (62–64), such as wearing sunglasses, a scarf or even a mask, is one critical factor that affects facial recognition performance (65,66). Moreover, according to Pham et al., distinguishing between emotions with similar facial expressions, such as fear and surprise, is complicated and causes low detection. However, a real-world scenario shows that dynamic environmental factors like facial obstacles, facial pose, and face and camera distance differences also cause low detection accuracy (20,26,65). Thus, as the main issue is low accuracy, tackling the issue of a dynamic environment with a general solution might help improve accuracy. Material and Method This section discusses the two types of datasets used and provides a brief explanation of our facial emotion detection framework. Materials We use two types of datasets in this study. The first dataset is the JAFFE dataset, a device-independent setting of facial expression images. The images were captured in a controlled environment, and the subjects were asked to mimic the expression. The second self-acquired dataset is Facial Expression Recognition 2013 (FER2013). This type of expression supports the real-world situation where no device setting can capture the expression. The dataset sample is shown in Figure 1. The JAFFE image set's first dataset included 213 images of ten Japanese female frontal facial photos, categorised into seven primary emotions (angry, disgust, fear, happy, neutral, sad and surprise) reflecting the system's classification (67,68). Each image in the JAFFE database has a resolution of 256x256 pixels. The images were taken under controlled conditions to guarantee consistency and reliability in the dataset. The JAFFEE database was designed and created by Michael Lyons, Miyuki Kamachi, and Jiro Gyoba (22,69,70). The JAFFE image set's first dataset included 213 images of ten Japanese female frontal facial photos, categorised into seven main emotions (angry, disgust, fear, happy, neutral, sad and surprise) reflecting the system's classification (67,68). Each image in the JAFFE database has a resolution of 256x256 pixels. The images were taken under controlled conditions to guarantee consistency and reliability in the dataset. The JAFFEE database was designed and created by Michael Lyons, Miyuki Kamachi, and Jiro Gyoba (22,69,70). While the second dataset is the FER2013 Dataset, as the study is focused on device-dependent setting facial expressions to imitate the real-world scenario, the latest version of Facial Expression Recognition 2013 (FER-2013) is used as facial data input. This expression exemplifies a scenario in the real world when there is no setup on any device that can capture that expression. The facial expression transition is more realistic, resembling the mimic expression captured in the laboratory (16). The dataset comprises grayscale images of faces with dimensions of 48x48 pixels. The facial features have been automatically aligned to ensure that the face is approximately centred and occupies an equivalent amount of space in each image. The objective is to classify each face according to the emotion shown in the facial expression into one of seven categories (0=Angry, 1=Disgust, 2=Fear, 3=Happy, 4=Sad, 5=Surprise, 6=Neutral). The training set has 28,709 samples, whereas the public test set comprises 3,589 examples. Method Overview The block diagram of the proposed framework is shown in Figure 2. The flowchart illustrates the process of facial emotion detection using two datasets: JAFFE and FER2013. The JAFFE dataset comprises facial images of Japanese female subjects displaying seven emotions, while the FER2013 dataset includes diverse subjects captured under various conditions, reflecting common issues in real-world images. The process starts with the JAFFE dataset serving as the training dataset. Feature extraction uses Wavelet Coefficient (WC) to extract meaningful features from each image. Wavelet Coefficients are beneficial as they effectively capture spatial and frequency information in the images, providing a robust representation of facial features crucial for accurate emotion detection. WC is good at resisting noises in images as no pre-processing is done. Moreover, WCs can reflect changes in grayscale images (71), and to solve the illumination issues, WCs are less sensitive to variation in lighting, which helps ensure significant features are extracted (72). The extracted features are then fed into a Support Vector Machine (SVM) to classify the images into seven emotions: Angry, Happy, Disgust, Sad, Fear, Surprise, and Neutral. SVM is one of the most used classifiers and can achieve high accuracy (71). SVM does perform well in terms of dimensionality and memory management (73). The extracted features are then fed into a Support Vector Machine (SVM) to classify the images into seven emotions: Angry, Happy, Disgust, Sad, Fear, Surprise, and Neutral. SVM is one of the most commonly used classifiers and can achieve high accuracy (71). SVM performs well in terms of dimensionality and memory management (73). The framework is subsequently tested using the FER2013 dataset, miming real-world situations with diverse subjects and varying conditions. This step assesses the effectiveness of the trained model in practical applications. The results and implications are discussed in the next section, evaluating the model's performance in a more challenging and realistic setting. Result and Discussion This section provides a comprehensive analysis of the results obtained from the experiments performed throughout the testing phase of the facial emotion detection framework. Comparison of histogram intensities for seven emotions of JAFFE. Image histograms are frequently used throughout various image processing applications, including image enhancement, segmentation, and analysis (27–29). Histograms are graphic representations that illustrate the frequency of pixel occurrences for each intensity value in a picture, thereby displaying the distribution of pixel intensities. An 8-bit grayscale image has 256 distinct intensities, ranging from 0 to 255. These values facilitate an in-depth understanding of the image's features, wherein several peaks may suggest the presence of distinct objects or regions in the image or offer insights into the general illumination, differences, and range of the image. The graph displays intensity values on the x-axis and the frequency or number of pixels at each intensity value on the y-axis (30). Additionally, it can perform advanced analysis such as histogram equalisation, histogram specification, and local enhancement. Figure 3 shows an example of histogram intensity analysis done towards the JAFFE dataset. For the Angry emotion, all three graphs exhibit a significant peak in the region of pixel value 50, indicating a high frequency of dark pixels. These obscure particles could indicate occluded areas or concealed facial features. At mid-range intensity, all three graphs exhibit several peaks. The frequency of pixel occurrences in the third graph is the lowest compared to the other two. Upon comparing the images, the facial expression displays lower signs of anger and less tension in the lips. These variations in facial features may result in distinct shadows, therefore implying a different underlying reason for the shape of the graph. On the other hand, the right side of the graph displays a uniform distribution of bright pixels throughout all three graphs. Meanwhile, the dark pixel region across the disgust emotion analysis exhibits three distinct frequency occurrences. The first graph displays the highest number of peaks in this pattern, indicating a correlation with obscured facial features or shadowed areas. This suggests that the first set of images portrays more emotions of disgust compared to the other two sets of images. For mid-range intensities, the graph distribution was similar. However, there was a noticeable variation in the peak observed for image number one, indicating that this image displayed the highest level of severe disgust emotion compared to the other two images. Finally, the bright pixels persist but are reduced in all three graphs. Meanwhile, the histogram graph shapes of the three images for the Fear emotion at mid-range intensities differ. In the first graph, the histogram displays a trough where fewer pixels are found within this range, indicating a lower degree of variability in the features. The second image demonstrates a more uniform distribution of pixels. This suggests that the characteristics within this range differ, indicating a less intense expression. Regarding the third image, there is a constant distribution of mid-range intensities, indicating a balanced representation of facial features throughout this range. Meanwhile, the distribution of graph values for dark and bright pixels in these three graphs is subsequently the same. In the context of low intensities, the histogram exhibits a peak around a pixel value of 50, while for high intensities, the pixels are still present but less obvious. Three images portraying the emotion of happy from the JAFFE dataset exhibit distinct styles of expression. The first and third images show the subject smiling with visible teeth. However, the second image does not display this characteristic. Regarding mid-range intensities, minor variations exist. The initial graph displays a histogram with a valley, indicating a reduced feature variation and a smaller number of pixels falling within this range. The second and third graphs exhibit a more uniform and evenly distributed range of mid-level intensities. The pixel intensities of all three images demonstrate a consistent pattern, with the histogram displaying a flattened peak and suppressed pixel values starting from 200 onwards. Each graph displays a single significant peak with a pixel value of roughly 50 for the dark intensities of the Neutral emotion histogram. All graphs display evenly distributed pixels, indicating a neutral expression. The histograms for mid-range intensities displayed uniform distribution across all graphs, suggesting a more subdued expression like neutrality. Finally, the bright intensities in all graphs are still visible but toned down, indicating neutrality. These illuminated areas suggest a neutral image. The dark pixels in all three sad images peak at a pixel value of 50 on the left side of the graphs. For mid-range intensities in the first graph, the middle section of the histogram displays a valley where fewer pixels are present, indicating reduced feature variability. Image two displays a more uniform distribution of pixel intensities, while the last image also shows an identical distribution, indicating balanced facial features within this range. For bright intensities around 200, the pixel intensities are visible but not as noticeable in all graphs. Finally, the histograms of surprise expressions each peak at approximately pixel value 50, indicating low intensities of dark pixels. The dark pixels may represent obscured features of the face or shaded areas caused by wrinkles on the forehead during a surprised expression. There are fewer pixels in this range for mid-range intensities in image number one, indicating less variability in features compared to images number two and number three. The mid-range intensities exhibit a uniform distribution. For bright pixels in all images, the pixels are evenly distributed, suggesting a predictable emotional state. Distinct patterns are revealed by a closer look at histograms representing emotions like anger, disgust, fear, happy, neutrality, sad, and surprise. The patterns, with peaks and valleys in the histograms, represent changes in facial expressions and emotional intensity. Anger usually peaks at low intensities, but happy tends to have more consistent distributions. Furthermore, histograms provide vital insights into the intricate features of many emotional states, aiding in advancing emotion detection and understanding. However, this study found that angry emotion displays the most significant changes in pixel intensities, as shown by dark and mid-range regions compared to other emotions. The analysis of histograms provides an in-depth understanding of facial expressions and associated emotions. Studying the distribution of pixel intensities allows us to uncover essential details about human emotions, leading to progress in emotion detection and understanding. To gain a deeper understanding of how different facial regions can give different meanings of expression for various emotions, the facial images were segmented into many areas of interest (ROI). The eyebrows, eyes, and mouth were selected as significant ROI. The pixel intensities are shown in Figure 4. From this experiment, the distribution of pixel intensities according to the selected ROI has unique and significant meanings across different emotions. More dark pixels in the eyebrow region may indicate an angry or sad emotion. The following experiment will confirm the detailed analysis. As previously discussed, angry emotions show significant changes in pixel intensities in the dark and middle-range regions. Thus, the image segmentation into three areas of interest revealed detailed differences for each emotion. The study found that the angry emotion had a high frequency of dark pixels, indicating significant shading due to furrowed eyebrows compared to other emotions. disgust, fear, and Happy also exhibit high intensities in the dark region, suggesting that these emotions are associated with eyebrow movements, such as furrowing or raising. In contrast with the surprise emotion, the graph skewed to the left, indicating an eyebrow movement that results in darker pixels. Sad and surprise are emotions typically closely associated with the eyes. When we compared the histogram of pixel intensities for all emotions, two distinct types of graphs emerged. First is the graph, which has a concentrated shading of dark pixels, stable mid-range intensities, and even a distribution of bright pixels, which brings angry, disgust, happy, sad, and neutral emotions. Fear and neutral show another graph pattern with less variable mid-range intensities, indicating consistent shading and evenly distributed bright pixels. This suggests that the eyes show changes only with specific emotions, suggesting that the eyes are more significant, smaller, or have reflection due to the tears. Finally, the third ROI shows something different, with graphs showing three types of shapes. The first graph shows a left-skewed pattern, signifying a highly concentrated dark region. The second graph displays a consistent pixel distribution across all three sections (dark, middle and lighter pixels), and the final graph displays a right-skewed pattern. Given the visual significance of the mouth movement associated with surprise, this study shows that the skewed right graph corresponds to the emotion of surprise. The skewed right graph indicates a slight movement of the mouth in line with the emotions of anger, fear, and neutrality. However, disgust, Happy, and sad exhibit an equal distribution of pixels across all types of shading, indicating more significant mouth movement. Emotion Classification Using SVM on the JAFFE Dataset This section describes the experimental results of SVM performances on the JAFFE dataset. Emotion recognition system sensitivity is the ability to precisely identify specific emotions. Following this measurement, the system is evaluated for its ability to accurately identify actual positive cases, meaning situations where the system reliably determines the existence of a particular emotion. It assesses how correctly the algorithm captures the actual instances of the emotion. High sensitivity in this context refers to the ability of an emotion detection system to accurately identify true expressions of Happy, sad, anger, and other emotions. Conversely, poor sensitivity suggests that the system may overlook certain instances of these emotions. Specificity is the ability of an emotion detection system to accurately identify persons who do not display a particular emotion. The system is evaluated on its capacity to identify true negative cases, namely when it accurately detects the lack of emotion. It measures the system's ability to accurately classify occurrences that are not emotional, thus avoiding false positives. A highly specialised emotion detection system can accurately identify when a person is not displaying anger, sad, or any other specified emotion. On the other hand, a low specificity rate suggests that the system might inaccurately categorise non-emotional occurrences as the correct emotion. Table 2 shows previous researchers' misclassification rate on JAFEE compared to our framework. Abbreviations represent the emotions: AN (Anger), DI (Disgust), FE (Fear), HA (Happy), NE (Neutral), SA (Sad), and SU. (Surprise). YOLO method shows higher misclassification rates overall, with the highest for surprise (39.0%) and the lowest for Happy (11.0%). Meanwhile, Stacked Sparse Auto-Encoder shows the lowest misclassification rates, with a good percentage of misclassification rate for most emotions, particularly surprise (1.0%) and Fear (1.0%). Our method shows a significantly higher misclassification rate for Happy, which is 33.0% and the lowest rate for Neutral emotion (1.0%). This is supported by the lowest result of specificity of Happy emotion from Table 3, 61.82% and the lowest result of accuracy, 67.19%. When sensitivity is low, causing the system to fail to identify a significant number of true positive emotions. As a result, individuals who genuinely express an emotion are missed by the emotion detection model. High false negatives occur when true positive instances are not correctly detected. The accuracy rates show how well an algorithm or system categorises different emotions. The method used accurately detects emotions of anger 92.19% of the time. The accuracy rate climbs to 96.88% when recognising disgust-related expressions, demonstrating an excellent capability in this domain. The model effectively manages the emotion of fear, reaching an accuracy rate of 93.75%. The accuracy decreases significantly to 67.19% while dealing with happy emotions. The model performs exceptionally well at identifying neutral emotions, achieving an accuracy rate of 98.44%. Sad, anger, and fear are categorised with an accuracy of 92.19%. Surprise emotions are identified with 93.75% accuracy. The accuracy rates offer essential insights into the model's performance across multiple emotional categories. Achieving consistently high accuracy across all emotions can be challenging due to the complexity and subjectivity associated with human emotions. The specificity rates for different emotions offer insights into the model's accuracy in distinguishing between various emotional states. The specificity rate for classifying anger is 100%, demonstrating a high level of accuracy in detecting anger, disgust, fear, neutrality, sad, and surprise without misclassifying other emotions. The specificity rate for happy is 61.82%, indicating that the proposed method has incorrectly categorised other emotions as Happy as the specificity rates provide essential information about the method's capacity to minimise false positives and correctly categorise various emotional states. The rates are impacted by things including training data and model architecture. Table 3. Misclassification rate on JAFFE by previous researchers Misclassification rate (%) Classification AN DI FE HA. NE SA. SU. Yolo (44) 19.0 12.0 15.0 11.0 13.0 19.0 39.0 CNN (41) 2.0 6.0 8.0 4.0 5.0 2.0 4.0 Stacked Sparse Auto-Encoder (45) 2.0 3.0 1.0 1.0 5.0 3.0 1.0 SVM (This study) 8.0 4.0 7.0 33.0 2.0 8.0 7.0 Table 4. SVM classification performance measurement on JAFFE Emotion Accuracy (%) Sensitivity (%) Specificity (%) Angry 92.19 37.50 100 Disgust 96.88 75.00 100 Fear 93.75 63.64 100 Happy 67.19 100.00 61.82 Neutral 98.44 85.71 100 Sad 92.19 64.29 100 Surprise 93.75 42.86 100 Analysis of emotion detection on JAFFE This section discussed performance results on the JAFFE dataset by comparing two methods. The first method uses only Support Vector Machines (SVM) in the framework, and the second method uses an additional Wavelet Coefficient together with SVM (WC+SVM). This experiment aims to see whether WC can enhance the classification performance for each emotion or not. Table 5. Comparison of accuracy according to emotions Emotion Accuracy of SVM (%) Accuracy of WC+SVM (%) Angry 92.19 96.88 Disgust 96.88 95.31 Fear 93.75 95.31 Happy 67.19 95.31 Neutral 98.44 96.88 Sad 92.19 84.38 Surprise 93.75 95.31 Average 90.10 94.20 Significant differences in performance are observed when comparing SVM with WC+SVM across different emotional classes. WC+SVM yields a 96.88% accuracy rate in anger emotion, showing an enhancement compared to 92.19% accuracy using SVM. WC+SVM maintains a consistent accuracy rate of 95.31% for disgust and fear emotion, while the accuracy rate for SVM varies slightly around 96.88% and 93.75% for the respective emotions. There is a significant increment in the accuracy of happy emotion when using WC+SVM, achieving 95.31% accuracy compared to 67.19% only using SVM. For the accuracy rate of classification for sad emotion, WC+SVM has an accuracy rate of 84.3ST8%, whereas SVM retains a better accuracy rate of 92.19%. WC+SVM achieves accuracy rates like SVM for neutral and surprise emotions, with minor variations but no significant differences. WC+SVM obtains an accuracy of 96.88% for neutral emotion and 95.31% for surprise, slightly surpassing SVM's accuracy. WC+SVM outperforms SVM with an average accuracy of 94.20% compared to SVM average accuracy of 90.10%. The results indicate that using wavelet coefficient as feature extraction improves emotion detection accuracy, especially in identifying happy emotions. However, it may have a little lower accuracy for sad emotions. The Wavelet Coefficient technique is excellent at capturing spatial and spectral information. This ability is beneficial for identifying significant features in facial expressions, allowing for a more detailed investigation. Furthermore, this method is very efficient for individuals dealing with large datasets and aiming to extract thorough features. Cross-database evaluation performance Previous researchers frequently use a variety of well-known, publicly accessible facial expression databases to report on the accuracy of the suggested methodologies (74). Previous studies have demonstrated successful detection performance on particular datasets under intra-dataset testing scenarios (75). According to Wen et al., a trained classifier might perform well on a specific set of training samples, but it may not be as effective in real-world scenarios where the sample size is much larger. Furthermore, real-world data usually causes a decrease in accuracy (39,76). This is due to the lab-dependent setting, such as the expression, lighting and posture of the subject by the public facial databases (43). In the deep learning age, large-scale, high-quality datasets are an especially critical prerequisite for facial expression recognition (FER); nevertheless, most FER datasets are quite modest. The cross-datasets strategy is a popular approach to solving this issue (77). Previous research publications fail to conduct cross-database evaluations (32,74,78). Thus, it is important to have a generalisation process where the classifier can perform well on unseen data and be more practical and valuable. Nevertheless, only a few research studies do cross-dataset testing evaluations to demonstrate the reliability of facial expressions identified by unfamiliar sources. Hybrid approaches combining several descriptors, face regions, picture modalities, and visual cues, including head movement and body gestures, need more research to enhance the effectiveness of facial expression identification. In the deep learning age, large-scale, high-quality datasets are an especially critical prerequisite for facial expression recognition (FER); nevertheless, most FER datasets are quite modest. The cross-datasets strategy is a popular approach to solving this issue (77). Previous research publications fail to conduct cross-database evaluations (32,74,78). Thus, it is important to have a generalisation process where the classifier can perform well on unseen data and be more practical and valuable. Nevertheless, only a few research studies do cross-dataset testing evaluations to demonstrate the reliability of facial expressions identified by unfamiliar sources. Hybrid approaches combining several descriptors, face regions, picture modalities, and visual cues, including head movement and body gestures, need more research to enhance the effectiveness of facial expression identification. Nowadays, most FER methods are developed and assessed with the presumption that the dataset containing the training and testing facial expression images will be the same (76). A periodic facial expression recognition challenge might offer a reasonable basis for this comparison. It would provide a clearer picture of the advancement of the field and enable us to establish new objectives, obstacles, and targets. Table 5 compares the numbers of different methods for facial emotion detection, outlining different classification methods, different sets of training and testing datasets and average accuracy achieved. The table gives a general overview of the several techniques used for emotion intensity analysis, including ECNN, CNN, Gabor with hierarchical classifier, SVM, RF and SVM, Local Directional Pattern with SVM, Boosted-LBP, Transfer Learning with SVM, and Wavelet Coefficient with SVM. Most authors used seven common types of emotion except Mayer (79), where the research only used six basic emotions that excluded neutral emotion. There are noticeable differences in accuracy and performance between the given approaches for various datasets. For example, the ECNN approach obtained an accuracy of 76.05% when evaluated on CK+ and trained on the FER2013 dataset (32). Comparably, the CNN approach achieved an accuracy of 73.38% after being tested on CK+ but trained on the same dataset (32). Nevertheless, the accuracy of these techniques significantly dropped when they were used on the JAFFE dataset, indicating possible difficulties in generalising across various datasets. In contrast to the CNN and ECNN approaches, the Gabor with hierarchical classifier method, evaluated on CK+ and trained on the JAFFE dataset, they demonstrated an acceptable performance standard with an accuracy of 54.05% Gu et al., 2012). Conversely, the accuracies of SVM-based techniques, such as those by Mayer et al., varied according to the training dataset used. For example, SVM obtained an accuracy of 66.20% when evaluated on CK+ and trained on the MMI dataset (79). After training on the BU-3DFE dataset and testing on JAFFE, an SVM-based method achieved an accuracy of 41.96%, suggesting lower accuracy on this dataset. Zhu et al. suggested Transfer Learning with SVM showed varying accuracies depending on whether the data was normalised (54). The normalised strategy achieved an accuracy of 60.09% when tested on JAFFE and trained on the FEED dataset, while the non-normalised approach achieved an accuracy of 46.48%. Specifically, our suggested approach, Wavelet Coefficient with SVM, showed encouraging outcomes, averaging an extraordinary 78.29% accuracy. Our solution outperformed numerous other methods in the table, demonstrating robust performance after training on the JAFFE dataset and testing on an occluded subset of FER2013. This indicates the effectiveness of using SVM classification in conjunction with wavelet coefficient analysis for tasks related to emotion intensity analysis. Table 6. Performance comparison between our framework and previous works that use cross-database method Method Authors Training Dataset Testing Dataset Average Accuracy Wavelet Coefficient with SVM This study JAFFE Occluded FER2013 78.29 ECNN (32) FER2013 CK+ 76.05 CNN (32) FER2013 CK+ 73.38 Gabor with hierarchical classifier (80) JAFFE CK+ 54.05 SVM (79) MMI CK+ 66.20 SVM (79) FEED CK+ 56.60 ECNN (32) FER2013 JAFFE 50.70 CNN (32) FER2013 JAFFE 45.07 Gabor with hierarchical classifier (80) CK+ JAFFE 55.87 RF and SVM (81) BU-3DFE JAFFE 41.96 Local Directional Pattern with SVM (57) CK JAFFE 45.71 Boosted-LBP (82) CK JAFFE 41.30 Transfer Learning with SVM (54) FEED JAFFE 46.48 Transfer Learning with SVM (54) FEED JAFFE 60.09 Our research demonstrates how well our proposed method can identify different facial emotions. We trained our model using the JAFFE dataset. Then, we tested it on obscured data from the FER2013 dataset, an excellent representation of the data we’d encounter in everyday life, as it closely resembles real-world scenarios. The key takeaway from the table is the importance of considering factors like the methodology used and variations in the datasets when evaluating the performance of emotion intensity analysis methods. This study needs deeper research and development in this area. We need to understand that specific methods are only suitable for certain situations or issues; thus, there is a need to have a technique that can generally overcome all the problems simultaneously. Conclusion Our research demonstrates that wavelet coefficient analysis with Support Vector Machine (SVM) classification provides a scalable and effective solution for detecting emotion intensities, particularly within large datasets. This framework’s robustness in handling occlusions makes it well-suited for real-time applications, such as video surveillance, where faces may be partially obscured, and real-time data processing is essential. To assess detection accuracy, we combined the wavelet coefficient with a cross-database method, testing its performance across various datasets, including the JAFFE dataset, which contains seven emotions, and FER2013, which closely mimics real-world conditions. We examined the impact of pixel density adjustments before applying wavelet analysis to optimise detection accuracy, showing that the framework consistently outperforms existing methods. However, it has yet to achieve 100% accuracy. Our approach shows strong potential for applications in psychological studies and social media emotion analysis, where scalable data processing and adaptability are critical. Experimental results confirm the framework’s ability to navigate the intricacies of emotion detection across varied, occluded datasets, supporting its viability in big data-driven environments. To further refine the accuracy of emotion representation, we plan to integrate additional feature extraction and selection methods in future studies, along with testing on more wild, real-world databases. This will ensure that the framework remains robust across diverse data types and comprehensively addresses current facial emotion detection challenges. Abbreviation Support Vector Machine (SVM), Wavelet Coefficient (WC), Japanese Female Facial Expression (JAFFE), Facial Expression Recognition 2013 (FER2013), AN (Anger), DI (Disgust), FE (Fear), HA (Happy), NE (Neutral), SA (Sad), and SU. (Surprise) Declarations Authors' contributions ASA designed and wrote the manuscript. RH contributed to the design and critically reviewed the manuscript. NHZ, SHM, ARK, MLMS, and SK participated in reviewing the manuscript to enhance its clarity and coherence. All authors read and approved the final manuscript. Acknowledgements This work was supported by the Ministry of Higher Education, Malaysia (MOHE) under the Fundamental Research Grant Scheme (FRGS), FRGS/1/2023/ICT02/UTM/02/8. References Kumar A, Kaur A, Kumar M. Face detection techniques: a review. Artif Intell Rev [Internet]. 2019 Aug 15 [cited 2020 Nov 30];52(2):927–48. Available from: https://doi.org/10.1007/s10462-018-9650-2 Revina IM, Emmanuel WRS. A Survey on Human Face Expression Recognition Techniques. J King Saud Univ - Comput Inf Sci [Internet]. 2018; Available from: https://doi.org/10.1016/j.jksuci.2018.09.002 Cowen A, Sauter D, Tracy JL, Keltner D. Mapping the Passions: Toward a High-Dimensional Taxonomy of Emotional Experience and Expression. Psychol Sci Public Interes [Internet]. 2019 Jul 17 [cited 2019 Oct 15];20(1):69–90. Available from: http://journals.sagepub.com/doi/10.1177/1529100619850176 Ekman P. Facial expression and emotion. Am Psychol [Internet]. 1993 [cited 2020 Nov 25];48(4):384–92. Available from: /record/1993-32252-001 Ekman P. Facial expressions of emotion: an old controversy and new findings. [Internet]. Vol. 335, Philosophical transactions of the Royal Society of London. Series B, Biological sciences. Philos Trans R Soc Lond B Biol Sci; 1992 [cited 2020 Nov 25]. p. 63–9. Available from: https://pubmed.ncbi.nlm.nih.gov/1348139/ Ekman P, Freisen W V., Ancoli S. Facial signs of emotional experience. J Pers Soc Psychol [Internet]. 1980 Dec [cited 2020 Nov 25];39(6):1125–34. Available from: /record/1981-25797-001 Bahreini K, van der Vegt W, Westera W. A fuzzy logic approach to reliable real-time recognition of facial emotions. Multimed Tools Appl [Internet]. 2019 Jul 30 [cited 2020 Mar 6];78(14):18943–66. Available from: https://doi.org/10.1007/s11042-019-7250-z Dolan E, Hancock E, Wareing A. An evaluation of online learning to teach practical competencies in undergraduate health science students. Internet High Educ. 2015 Jan 1;24:21–5. Kumari J, Rajesh R, Pooja KM. Facial Expression Recognition: A Survey. Procedia Comput Sci [Internet]. 2015 Jan 1 [cited 2020 Dec 3];58:486–91. Available from: http://dx.doi.org/10.1016/j.procs.2015.08.011 Pantic M, Valstar M, Rademaker R, Maat L. Web-based database for facial expression analysis. In: IEEE International Conference on Multimedia and Expo, ICME 2005. 2005. p. 317–21. Yin L, Wei X, Sun Y, Wang J, Rosato MJ. A 3D facial expression database for facial behavior research. In: FGR 2006: Proceedings of the 7th International Conference on Automatic Face and Gesture Recognition. 2006. p. 211–6. Gross R, Matthews I, Cohn J, Kanade T, Baker S. Multi-PIE. Image Vis Comput. 2010 May 1;28(5):807–13. Lucey P, Cohn JF, Kanade T, Saragih J, Ambadar Z, Matthews I. The extended Cohn-Kanade dataset (CK+): A complete dataset for action unit and emotion-specified expression. In: 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Workshops, CVPRW 2010. 2010. p. 94–101. Langner O, Dotsch R, Bijlstra G, Wigboldus DHJ, Hawk ST, van Knippenberg A. Presentation and validation of the radboud faces database. Cogn Emot [Internet]. 2010 Dec [cited 2020 Dec 3];24(8):1377–88. Available from: https://www.tandfonline.com/doi/abs/10.1080/02699930903485076 Valstar MF, Jiang B, Mehu M, Pantic M, Scherer K. The first facial expression recognition and analysis challenge. In: 2011 IEEE International Conference on Automatic Face and Gesture Recognition and Workshops, FG 2011. 2011. p. 921–6. Dhall A, Goecke R, Lucey S, Gedeon T. Static facial expression analysis in tough conditions: Data, evaluation protocol and benchmark. In: Proceedings of the IEEE International Conference on Computer Vision. 2011. p. 2106–12. Zhao G, Huang X, Taini M, Li SZ, Pietikäinen M. Facial expression recognition from near-infrared videos. Image Vis Comput. 2011 Aug 1;29(9):607–19. Goodfellow IJ, Erhan D, Carrier PL, Courville A, Mirza M, Hamner B, et al. Challenges in representation learning: A report on three machine learning contests. In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) [Internet]. Springer, Berlin, Heidelberg; 2013 [cited 2020 Dec 3]. p. 117–24. Available from: http://www.kaggle.com/c/challenges-in-representation-learning-facial- Yan W-J, Li X, Wang S-J, Zhao G, Liu Y-J, Chen Y-H, et al. CASME II: An Improved Spontaneous Micro-Expression Database and the Baseline Evaluation. Guo K, editor. PLoS One [Internet]. 2014 Jan 27 [cited 2020 Dec 3];9(1):e86041. Available from: https://dx.plos.org/10.1371/journal.pone.0086041 Li S, Deng W, Du JP. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In: Proceedings - 30th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017. Institute of Electrical and Electronics Engineers Inc.; 2017. p. 2584–93. Mollahosseini A, Hasani B, Mahoor MH. AffectNet: A Database for Facial Expression, Valence, and Arousal Computing in the Wild. IEEE Trans Affect Comput. 2019 Jan 1;10(1):18–31. Lyons MJ, Kamachi M, Gyoba J. Coding Facial Expressions with Gabor Wavelets (IVC Special Issue). 2020 Sep 13 [cited 2020 Dec 3]; Available from: http://arxiv.org/abs/2009.05938 Poulose A, Kim JH, Han DS. Feature Vector Extraction Technique for Facial Emotion Recognition Using Facial Landmarks. Int Conf ICT Converg. 2021;2021-October:1072–6. Kishan Kondaveeti H, Vishal Goud M. Emotion Detection using Deep Facial Features. Proc IEEE Int Conf Advent Trends Multidiscip Res Innov ICATMRI 2020. 2020 Dec 30; Singh S, Prasad SVAV. Techniques and challenges of face recognition: A critical review. In: Procedia Computer Science. Elsevier BV; 2018. p. 536–43. Pham TTD, Kim S, Lu Y, Jung SW, Won CS Facial Action Units-Based Image Retrieval for Facial Expression Recognition. IEEE Access. 2019;7:5200–7. Trongtirakul T, Phanthuna N. Image Enhancement Using Weighted Bi-Histogram Equalization. Int J Appl Math Informatics. 2021;15:98–101. Barik R, Saha T, Chatterjee S. Reduction based histogram equalisation technique for image enhancement. Int J Innov Technol Explor Eng [Internet]. 2019 [cited 2024 Feb 15];8(11 Special Issue):1268–71. Available from: http://www.mathworks.com/products/demos/# Jumiawi WAH, El-Zaart A. Improvement in the Between-Class Variance Based on Lognormal Distribution for Accurate Image Segmentation. Entropy [Internet]. 2022 Aug 29 [cited 2024 Feb 15];24(9):1204. Available from: https://www.mdpi.com/1099-4300/24/9/1204/htm Sysko LR, Davis MA. From image to data using common image-processing techniques. Curr Protoc Cytom [Internet]. 2010 Oct 1 [cited 2024 Feb 15];54(SUPPL.54). Available from: https://typeset.io/papers/from-image-to-data-using-common-image-processing-techniques-2pwtgq1yha Liliana DY, Basaruddin T, Widyanto MR, Oriza IID. Fuzzy emotion: a natural approach to automatic facial expression recognition from psychological perspective using fuzzy system. Cogn Process [Internet]. 2019 Nov 1 [cited 2020 Mar 6];20(4):391–403. Available from: http://www.ncbi.nlm.nih.gov/pubmed/31209637 Wen G, Hou Z, Li H, Li D, Jiang L, Xun E. Ensemble of Deep Neural Networks with Probability-Based Fusion for Facial Expression Recognition. Cognit Comput. 2017 Oct 1;9(5):597–610. Shen T, Xu H. Facial Expression Recognition Based onMulti-Channel Attention Residual Network. C - Comput Model Eng Sci [Internet]. 2023 Sep 29 [cited 2024 Mar 14];135(1):539–60. Available from: https://www.techscience.com/CMES/v135n1/50070/html Appasaheb Borgalli MR, Surve D.S. Deep learning for facial emotion recognition using custom CNN architecture. In: Journal of Physics: Conference Series. IOP Publishing; 2022. p. 12004. An Y, Lee J, Bak ES, Pan S. Deep Facial Emotion Recognition Using Local Features Based on Facial Landmarks for Security System. Comput Mater Contin [Internet]. 2023 Aug 30 [cited 2023 Dec 8];76(2):1817–32. Available from: https://www.techscience.com/cmc/v76n2/54000/html Nicolai A, Choi A. Facial Emotion Recognition Using Fuzzy Systems. In: Proceedings - 2015 IEEE International Conference on Systems, Man, and Cybernetics, SMC 2015. Institute of Electrical and Electronics Engineers Inc.; 2016. p. 2216–21. Kim JH, Kim BG, Roy PP, Jeong DM. Efficient facial expression recognition algorithm based on hierarchical deep neural network structure. IEEE Access. 2019;7:41273–85. Zaman K, Zhaoyun S, Shah B, Hussain T, Shah SM, Ali F, et al. A novel driver emotion recognition system based on deep ensemble classification. Complex Intell Syst [Internet]. 2023 [cited 2023 Dec 25];9(6):6927–52. Available from: https://doi.org/10.1007/s40747-023-01100-9 Ramis S, Buades JM, Perales FJ, Manresa-Yee C. A Novel Approach to Cross dataset studies in Facial Expression Recognition. Multimed Tools Appl [Internet]. 2022 Nov 1 [cited 2024 Feb 29];81(27):39507–44. Available from: https://link.springer.com/article/10.1007/s11042-022-13117-2 Wang Y, Li Y, Song Y, Rong X. Facial expression recognition based on auxiliary models. Algorithms [Internet]. 2019 Oct 31 [cited 2019 Dec 7];12(11):227. Available from: https://www.mdpi.com/1999-4893/12/11/227 Benisha S, Mirnalinee TT. Human Facial Emotion Recognition using Deep Neural Networks. Int Arab J Inf Technol [Internet]. 2023 [cited 2024 Mar 14];20(3):303–9. Available from: https://doi.org/10.34028/iajit/20/3/2 Podder T, Bhattacharya D, Majumder P, Balas VE. A feature boosted deep learning method for automatic facial expression recognition. PeerJ Comput Sci [Internet]. 2023 Jan 31 [cited 2024 Mar 14];9:e1216. Available from: https://peerj.com/articles/cs-1216 Chandra J, Annappa B, Rashmi Adyapady R. Cross-Database Facial Expression Recognition using CNN with Attention Mechanism. In: 2023 14th International Conference on Computing Communication and Networking Technologies, ICCCNT 2023. Institute of Electrical and Electronics Engineers Inc.; 2023. Zhang R, Ma R. Facial expression recognition method based on PSA—YOLO network. Front Neurorobot. 2023 Jan 17;16:1057983. Ahmad M, Saira, Alfandi O, Khattak AM, Qadri SF, Saeed IA, et al. Facial expression recognition using lightweight deep learning modeling. Math Biosci Eng [Internet]. 2023 [cited 2024 Mar 14];20(5):8208–25. Available from: http://www.aimspress.com/article/doi/10.3934/mbe.2023357 Akhand MAH, Roy S, Siddique N, Kamal MAS, Shimamura T. Facial emotion recognition using transfer learning in the deep CNN. Electron [Internet]. 2021 Apr 27 [cited 2022 Feb 16];10(9):1036. Available from: https://www.mdpi.com/2079-9292/10/9/1036/htm Chatterjee S, Shi H. A novel Neuro Fuzzy approach to human emotion determination. In: Proceedings - 2010 Digital Image Computing: Techniques and Applications, DICTA 2010. 2010. p. 282–7. Guo X, Lu S, Wang S, Lu Z, Zhang Y. DLSANet: Facial expression recognition with double-code LBP-layer spatial-attention network. IET Image Process [Internet]. 2023 Jul 1 [cited 2024 Mar 14];17(9):2659–72. Available from: https://onlinelibrary.wiley.com/doi/full/10.1049/ipr2.12817 Saeed VA. A Framework for Recognition of Facial Expression Using HOG Features. Int J Math Stat Comput Sci [Internet]. 2023 May 27 [cited 2024 Mar 14];2:1–8. Available from: https://ijmscs.org/index.php/ijmscs/article/view/7815 Shelke N, Upadhye S, Uparkar SS, Bawane H, Shivvanshi D, Kosekar S. Approaches for harvesting on emotion extraction from human facial expressions. Indian J Comput Sci Eng. 2021 Jul 1;12(4):921–44. Cornejo JYR, Pedrini H, Flórez-Revuelta F. Facial expression recognition with occlusions based on geometric representation. In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) [Internet]. Springer Verlag; 2015 [cited 2024 Mar 14]. p. 263–70. Available from: https://link.springer.com/chapter/10.1007/978-3-319-25751-8_32 Harakannanavar SS, Sapnakumari C, Ramachandra AC, Pramodhini R, Prashanth CR Performance Evaluation of Fusion Based Efficient Algorithm for Facial Expression Recognition. Indian J Sci Technol [Internet]. 2023 Jan 28 [cited 2024 Mar 14];16(4):266–76. Available from: https://indjst.org/articles/performance-evaluation-of-fusion-based-efficient-algorithm-for-facial-expression-recognition Subudhiray S, Palo HK, Das N. K-nearest neighbor based facial emotion recognition using effective features. IAES Int J Artif Intell [Internet]. 2023 Mar 1 [cited 2024 Mar 14];12(1):57. Available from: https://ijai.iaescore.com/index.php/IJAI/article/view/21156 Zhu R, Zhang T, Zhao Q, Wu Z. A transfer learning approach to cross-database facial expression recognition. In: Proceedings of 2015 International Conference on Biometrics, ICB 2015. Institute of Electrical and Electronics Engineers Inc.; 2015. p. 293–8. Liu Y, Li Y, Ma X, Song R. Facial expression recognition with fusion features extracted from salient facial areas. Sensors (Switzerland). 2017 Apr 1;17(4). Jeen Retna Kumar R, Sundaram M, Arumugam N, Kavitha V. Face feature extraction for emotion recognition using statistical parameters from subband selective multilevel stationary biorthogonal wavelet transform. Soft Comput [Internet]. 2021 Apr 1 [cited 2022 Sep 7];25(7):5483–501. Available from: https://dl.acm.org/doi/10.1007/s00500-020-05550-y Zhou J, Xu T, Gan J. Feature extraction based on local directional pattern with SVM decision-level fusion for facial expression recognition. Int J Bio-Science Bio-Technology. 2013;5(2):101–10. Lu Y, Wang S, Zhao W. Facial expression recognition based on discrete separable shearlet transform and feature selection. Algorithms [Internet]. 2019 Dec 31 [cited 2019 Dec 8];12(1):11. Available from: http://www.mdpi.com/1999-4893/12/1/11 Eng SK, Ali H, Cheah AY, Chong YF. Facial expression recognition in JAFFE and KDEF Datasets using histogram of oriented gradients and support vector machine. In: IOP Conference Series: Materials Science and Engineering [Internet]. IOP Publishing; 2019 [cited 2024 Mar 17]. p. 012031. Available from: https://iopscience.iop.org/article/10.1088/1757-899X/705/1/012031 Delazeri BR, Vera LL, Barddal JP, Koerich AL, Alceu De SB. Evaluation of Self-taught Learning-based Representations for Facial Emotion Recognition. In: Proceedings of the International Joint Conference on Neural Networks. 2022. Tuncer T, Dogan S, Subasi A. Automated facial expression recognition using novel textural transformation. J Ambient Intell Humaniz Comput [Internet]. 2023 Jul 1 [cited 2024 Mar 14];14(7):9435–49. Available from: https://link.springer.com/article/10.1007/s12652-023-04612-x Yang B, Wu J, Ikeda K, Hattori G, Sugano M, Iwasawa Y, et al. Face-mask-aware Facial Expression Recognition based on Face Parsing and Vision Transformer. Pattern Recognit Lett [Internet]. 2022 Dec 1 [cited 2024 Jun 18];164:173–82. Available from: https://dl.acm.org/doi/10.1016/j.patrec.2022.11.004 Saeed U. Facial micro-expressions as a soft biometric for person recognition. Pattern Recognit Lett [Internet]. 2021 Mar 1 [cited 2024 Jun 18];143:95–103. Available from: https://dl.acm.org/doi/10.1016/j.patrec.2020.12.021 Ahmad AS, Hassan R, Zakaria NH, Moi SH. Comparative studies of facial emotion detection in online learning. In: AIP Conference Proceedings [Internet]. American Institute of Physics Inc.; 2023 [cited 2024 Jan 24]. Available from: /aip/acp/article/2827/1/030009/2910737/Comparative-studies-of-facial-emotion-detection-in Canedo D, Neves AJR. Facial expression recognition using computer vision: A systematic review [Internet]. Vol. 9, Applied Sciences (Switzerland). MDPI AG; 2019 [cited 2019 Dec 7]. p. 4678. Available from: www.mdpi.com/journal/applsci Su Y, Yang Y, Guo Z, Yang W. Face recognition with occlusion. In: Proceedings - 3rd IAPR Asian Conference on Pattern Recognition, ACPR 2015. Institute of Electrical and Electronics Engineers Inc.; 2016. p. 670–4. Revina ImM, Emmanuel WRS, Sam Emmanuel W. A Survey on Human Face Expression Recognition Techniques. J King Saud Univ - Comput Inf Sci [Internet]. 2018 [cited 2023 Jun 14]; Available from: https://doi.org/10.1016/j.jksuci.2018.09.002 Kasim S, Hassan R, Zaini NHNH, Ahmad ASAS, Ramli AAAA, Saedudin RRRR. A Study on Facial Expression Recognition Using Local Binary Pattern. Int J Adv Sci Eng Inf Technol [Internet]. 2017 Oct 26 [cited 2019 May 29];7(5):1621–6. Available from: http://ijaseit.insightsociety.org/index.php?option=com_content&view=article&id=9&Itemid=1&article_id=3390 Putra WB, Arifin F. Real-Time Emotion Recognition System to Monitor Student’s Mood in a Classroom. In: Journal of Physics: Conference Series. 2019. Lyons M, Akamatsu S, Kamachi M, Gyoba J. Coding facial expressions with Gabor wavelets. In: Proceedings - 3rd IEEE International Conference on Automatic Face and Gesture Recognition, FG 1998. IEEE Computer Society; 1998. p. 200–5. Liu C, Wang A. Adaptive High-order Variation De-noising Method for Edge Detection with Wavelet Coefficients. KSII Trans Internet Inf Syst [Internet]. 2023 Feb 28 [cited 2024 Jun 13];17(2):412–34. Available from: https://link.springer.com/article/10.1007/s11432-006-0494-z Hu H. Illumination invariant face recognition based on dual-tree complex wavelet transform. IET Comput Vis. 2015 Apr 1;9(2):163–73. Qazi AS, Farooq MS, Rustam F, Villar MG, Rodríguez CL, Ashraf I. Emotion Detection Using Facial Expression Involving Occlusions and Tilt. Appl Sci [Internet]. 2022 Nov 20 [cited 2024 May 31];12(22):11797. Available from: https://www.mdpi.com/2076-3417/12/22/11797/htm Valstar MF, Mehu M, Jiang B, Pantic M, Scherer K. Meta-analysis of the first facial expression recognition challenge. IEEE Trans Syst Man, Cybern Part B Cybern [Internet]. 2012 [cited 2024 Feb 29];42(4):966–79. Available from: http://ieeexplore.ieee.org. Jia S, Wang S, Hu C, Webster PJ, Li X. Detection of Genuine and Posed Facial Expressions of Emotion: Databases and Methods [Internet]. Vol. 11, Frontiers in Psychology. Frontiers Media SA; 2021 [cited 2024 Feb 29]. p. 580287. Available from: www.frontiersin.org Xu X, Zheng W, Zong Y, Lu C, Jiang X. Sample Self-Revised Network for Cross-Dataset Facial Expression Recognition. In: Proceedings of the International Joint Conference on Neural Networks. Institute of Electrical and Electronics Engineers Inc.; 2022. Meng H, Yuan F, Tian Y, Yan T. Cross-datasets facial expression recognition via distance metric learning and teacher-student model. Multimed Tools Appl [Internet]. 2022 [cited 2024 Feb 29];81(4):5621–43. Available from: https://doi.org/10.1007/s11042-021-11765-4 Li S, Deng W. Deep Emotion Transfer Network for Cross-database Facial Expression Recognition. In: Proceedings - International Conference on Pattern Recognition. 2018. p. 3092–9. Mayer C, Eggers M, Radig B. Cross-database evaluation for facial expression recognition. Pattern Recognit Image Anal [Internet]. 2014 [cited 2024 Mar 7];24(1):124–32. Available from: https://sci-hub.se/10.1134/s1054661814010106 Gu W, Xiang C, Venkatesh Y V., Huang D, Lin H. Facial expression recognition using radial encoding of local Gabor features and classifier synthesis. Pattern Recognit. 2012 Jan 1;45(1):80–91. Abd El Meguid MK, Levine MD. Fully automated recognition of spontaneous facial expressions in videos using random forest classifiers. IEEE Trans Affect Comput. 2014;5(2):141–54. Shan C, Gong S, McOwan PW. Facial expression recognition based on Local Binary Patterns: A comprehensive study. Image Vis Comput [Internet]. 2009 May 4 [cited 2017 Aug 3];27(6):803–16. Available from: http://www.sciencedirect.com/science/article/pii/S0262885608001844 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5379043","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":374415915,"identity":"c49f9bf2-69a2-4ba6-b76f-2444667a8972","order_by":0,"name":"Asraful Syifaa' Ahmad","email":"data:image/png;base64,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","orcid":"","institution":"Universiti Teknologi Malaysia","correspondingAuthor":true,"prefix":"","firstName":"Asraful","middleName":"Syifaa'","lastName":"Ahmad","suffix":""},{"id":374415916,"identity":"e096610b-9330-4567-8351-e1c888396ab4","order_by":1,"name":"Rohayanti Hassan","email":"","orcid":"","institution":"Universiti Teknologi Malaysia","correspondingAuthor":false,"prefix":"","firstName":"Rohayanti","middleName":"","lastName":"Hassan","suffix":""},{"id":374415917,"identity":"4ef61470-bbcf-4fa2-b80f-8b4c4f7916b0","order_by":2,"name":"Noor Hidayah Zakaria","email":"","orcid":"","institution":"Universiti Teknologi Malaysia","correspondingAuthor":false,"prefix":"","firstName":"Noor","middleName":"Hidayah","lastName":"Zakaria","suffix":""},{"id":374415918,"identity":"477f11db-015e-4e11-bed5-977e2c83cda2","order_by":3,"name":"Hiew Moi Sim","email":"","orcid":"","institution":"Universiti Teknologi Malaysia","correspondingAuthor":false,"prefix":"","firstName":"Hiew","middleName":"Moi","lastName":"Sim","suffix":""},{"id":374415919,"identity":"7a46f545-e5da-4e2f-9308-c1cdd8af70d4","order_by":4,"name":"Alif Ridzuan Khairuddin","email":"","orcid":"","institution":"Universiti Teknologi Malaysia","correspondingAuthor":false,"prefix":"","firstName":"Alif","middleName":"Ridzuan","lastName":"Khairuddin","suffix":""},{"id":374415920,"identity":"0852a6ad-bf94-4a76-ad17-3e063e862aff","order_by":5,"name":"Muhammad Luqman Mohd Shafei","email":"","orcid":"","institution":"Universiti Teknologi Malaysia","correspondingAuthor":false,"prefix":"","firstName":"Muhammad","middleName":"Luqman Mohd","lastName":"Shafei","suffix":""},{"id":374415921,"identity":"fab8cca8-08c3-4437-9917-e5b9741613ec","order_by":6,"name":"Shahreen Kasim","email":"","orcid":"","institution":"Universiti Tun Hussein Onn Malaysia","correspondingAuthor":false,"prefix":"","firstName":"Shahreen","middleName":"","lastName":"Kasim","suffix":""}],"badges":[],"createdAt":"2024-11-02 15:53:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5379043/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5379043/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":69340417,"identity":"8a0b5bc2-fd6f-4a40-97fa-8af3e35ae443","added_by":"auto","created_at":"2024-11-19 10:56:00","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":194602,"visible":true,"origin":"","legend":"\u003cp\u003eSample images from JAFFE \u0026nbsp;\u0026nbsp;and FER2013 occluded dataset\u003c/p\u003e","description":"","filename":"Fig1.png","url":"https://assets-eu.researchsquare.com/files/rs-5379043/v1/0847d1ffd4e26a3c819c7c37.png"},{"id":69340419,"identity":"7642d2dc-392b-44b8-80b5-4ff147a9345b","added_by":"auto","created_at":"2024-11-19 10:56:00","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":147818,"visible":true,"origin":"","legend":"\u003cp\u003eFacial Emotion Detection Framework\u003c/p\u003e","description":"","filename":"Fig2.png","url":"https://assets-eu.researchsquare.com/files/rs-5379043/v1/45b60b0c448a92aacb4f9143.png"},{"id":69340750,"identity":"3c06c79c-fb2d-4d27-bbae-eb0bb7889bb2","added_by":"auto","created_at":"2024-11-19 11:04:00","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":225558,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of histogram \u0026nbsp;\u0026nbsp;intensities for seven emotions for the JAFFE database\u003c/p\u003e","description":"","filename":"Fig3.png","url":"https://assets-eu.researchsquare.com/files/rs-5379043/v1/538324d8bd1be4863be7b77b.png"},{"id":69340418,"identity":"4222df7c-7e2c-489d-b2fc-f5578cf06ce7","added_by":"auto","created_at":"2024-11-19 10:56:00","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":420603,"visible":true,"origin":"","legend":"\u003cp\u003eHistogram intensities of the JAFFE dataset according to ROI\u003c/p\u003e","description":"","filename":"Fig4.png","url":"https://assets-eu.researchsquare.com/files/rs-5379043/v1/49973d9de48dd8e31f56fef1.png"},{"id":91231319,"identity":"91e3677d-7c38-404c-b1af-4d0aceb7fcec","added_by":"auto","created_at":"2025-09-13 06:46:37","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2066613,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5379043/v1/66697900-da4e-4e8b-918b-93646bcef6b9.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Spontaneous vs Posed: Analysis of Emotion Intensities using Wavelet Coefficient and Support Vector Machine","fulltext":[{"header":"Introduction","content":"\u003cp\u003eEmotion detection is a process of recognising and categorising human emotion from any source, including verbal and non-verbal communication. Facial expression and body language, including body posture and gesture, are non-verbal forms of communication\u0026nbsp;(1,2). Emotion represented by facial can be divided into six discrete categories\u0026nbsp;(3,4). These six basic emotions are Happiness, sad, anger, fear, surprise, and disgust. Thus, previous researchers have investigated facial features vigorously for emotion analysis\u0026nbsp;(4\u0026ndash;6).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eEmotion detection has various applications, including mental health detection for conditions like depression, posttraumatic stress disorder (PTSD), and autism. It also benefits online education, allowing teachers to monitor and understand their students\u0026apos; emotions. In addition, the entertainment industry can utilise emotion detection to enhance user experiences, and e-commerce platforms can leverage it to gauge customer emotions. For instance, e-commerce platforms often provide rating and ranking systems where customers can express their emotions and satisfaction levels regarding the services they receive from providers. This information encoded in customer reviews can provide valuable insights into their emotions related to the purchase experience\u0026nbsp;(7\u0026ndash;9). Moreover, emotion detection has broad applications in diverse fields, such as real-time video surveillance systems, large-scale psychology studies, and social media sentiment analysis, where accurate identification of emotional states is crucial for interpreting human behaviour in various contexts.\u003c/p\u003e\n\u003cp\u003eFacial emotion detection and facial expression recognition are interconnected areas that involve the analysis of facial expressions to understand a person\u0026apos;s emotions. While these terms are occasionally used interchangeably, facial emotion detection primarily focuses on recognising general emotional states, while facial expression recognition aims to identify specific expressions. Both fields utilise sophisticated algorithms and machine learning techniques to analyse facial features and find applications in mental health assessment and human-robot interaction.\u003c/p\u003e\n\u003cp\u003eHowever, despite technological advancements, facial emotion detection still faces significant challenges, such as occlusions, subtle expression variations, illumination differences, and low-resolution images, all affecting detection accuracy. These issues become even more critical when scaling detection methods to large datasets. Given these needs, our study proposes an improved facial emotion detection framework that leverages wavelet coefficient analysis and Support Vector Machine (SVM) classification to select the most significant features for each emotion, thereby enhancing detection accuracy. This approach addresses occlusion and other common obstacles and enhances system adaptability, making it suitable for high-performance applications in big data environments. To achieve this aim, this paper has conducted the following investigations:\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e1.\u0026nbsp; \u0026nbsp;Histogram Intensities Comparison Across Seven Emotions in the JAFFE Database.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e2.\u0026nbsp; \u0026nbsp;Emotion Classification Using Support Vector Machine on the JAFFE Dataset\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3. \u0026nbsp; Improvement in Accuracy by Having Wavelet Coefficient with Support Vector Machine\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e4. \u0026nbsp; Cross-Database Evaluation Performance Using JAFFE Dataset and Occluded FER2013 Dataset\u003c/p\u003e\n\u003cp\u003eThe remainder of this paper is organised as follows: Section 2 provides the related works according to emotion detection, and Section 3 provides an overview of the proposed method, techniques, and dataset for the whole framework. The results are presented and discussed in Section 4, while Section 5 concludes the paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRelated Works\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis section consists of an overview of public facial emotion datasets and recent feature extraction and classification techniques. The table provides an overview of various emotion detection datasets that have been publicly used. Each of the datasets is comprised of basic emotions captured from either posed participants, real-world situations, or video.\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;1. Dataset of the benchmark dataset\u003c/p\u003e\n\u003cdiv align=\"Left\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"605\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDataset\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 161px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eEmotion\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 217px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDescription\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 94px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSetting Image\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eMMI\u0026nbsp;(10)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eSix basic emotions \u0026amp; neutral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e2900 videos (the neutral, onset, apex and offset)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.5629%;\"\u003e\n \u003cp\u003ePosed\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eBU-3DFE\u0026nbsp;(11)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eSix basic emotions \u0026amp; neutral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e2500 3D facial images captured on two views -45\u0026deg;, +45\u0026deg;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.5629%;\"\u003e\n \u003cp\u003ePosed\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eMultiPie\u0026nbsp;(12)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eAnger, disgust, neutral, happy, squint, scream, surprise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e750,000 images (15 view and 19 illumination conditions)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.5629%;\"\u003e\n \u003cp\u003ePosed\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eCK+\u0026nbsp;(13)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eSix basic emotions, contempt \u0026amp; neutral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e593 videos (posed and non-posed expressions)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.5629%;\"\u003e\n \u003cp\u003eBoth\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eRaFD\u0026nbsp;(14)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eSix basic emotions, contempt \u0026amp; neutral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e8040 images (different face poses, ages, genders \u0026amp; sexes)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.5629%;\"\u003e\n \u003cp\u003ePosed\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eGEMEP FERA\u0026nbsp;(15)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eAnger, fear, sad, relief, happy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e289 images sequences\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.5629%;\"\u003e\n \u003cp\u003ePosed\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eSFEW\u0026nbsp;(16)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eSix basic emotions \u0026amp; neutral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e700 images (different ages, occlusion, illumination and head pose)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.5629%;\"\u003e\n \u003cp\u003eSpontaneous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eOulu-CASIA\u0026nbsp;(17)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eSix basic emotions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e2880 videos (three different illumination conditions)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.5629%;\"\u003e\n \u003cp\u003ePosed\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eFER2013\u0026nbsp;(18)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eSix basic emotions \u0026amp; neutral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e35,887 grayscale images (from Google Image search)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.5629%;\"\u003e\n \u003cp\u003eSpontaneous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eCASME II\u0026nbsp;(19)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eHappy, disgust, surprise, regression \u0026amp; others\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e247 micro-expressions sequences\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.5629%;\"\u003e\n \u003cp\u003ePosed\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eRAFD-DB\u0026nbsp;(20)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eSix basic emotions \u0026amp; neutral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e30,000 images from the real world\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.5629%;\"\u003e\n \u003cp\u003eSpontaneous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eAffectNet\u0026nbsp;(21)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eSix basic emotions \u0026amp; neutral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003eMore than 440,000 images collected from the internet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.5629%;\"\u003e\n \u003cp\u003ePosed\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.8543%;\"\u003e\n \u003cp\u003eJAFFE\u0026nbsp;(22)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.6556%;\"\u003e\n \u003cp\u003eSix basic emotions \u0026amp; neutral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e213 grayscale images posed by 10 Japanese females\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.5629%;\"\u003e\n \u003cp\u003ePosed\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eThe recognition performance is affected by the number of features extracted as input for the classifiers\u0026nbsp;(23). Thus, a suitable extraction method is needed. Feature extraction is a process in recognition that extracts important and meaningful features that could represent facial images. Later, these features will be input to the classification step\u0026nbsp;(24).\u003c/p\u003e\n\u003cp\u003eFacial expressions hold abundant information regarding an individual\u0026apos;s internal thoughts and emotions, making them crucial for human communication. In real-world scenarios, the images captured contain odd expressions, illuminations, face occlusion, less resolution, different distances from the camera, and different angles of face orientation\u0026nbsp;(1,25,26).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eImage histograms are frequently used in various image processing applications, including image enhancement, segmentation, and analysis\u0026nbsp;(27\u0026ndash;29). Histograms are graphic representations that illustrate the frequency of pixel occurrences for each intensity value in a picture, thereby displaying the distribution of pixel intensities. An 8-bit grayscale image has 256 distinct intensities, ranging from 0 to 255. These values facilitate an in-depth understanding of the image\u0026apos;s features, wherein several peaks may suggest the presence of different objects or regions in the image or offer insights into the general illumination, differences, and range of the image. The graph displays intensity values on the x-axis and the frequency or number of pixels at each intensity value on the y-axis\u0026nbsp;(30). Additionally, it can perform advanced analysis such as histogram equalisation, histogram specification, and local enhancement.\u003c/p\u003e\n\u003cp\u003eThe table presents significant progress in emotion detection technology, featuring several approaches, including fuzzy inference systems, advanced neural network designs, and conventional supervised learning techniques. Over time, the accurateness of these methods has significantly enhanced, with modern techniques such as Transfer Learning + pre-trained DCNN and different Improved CNN models acquiring nearly perfect accuracy.\u003c/p\u003e\n\u003cp\u003eThe Fuzzy Emotion Inference System, created by Liliana et al. in 2019, employs a Fuzzy Inference Engine specifically built to manage uncertainty and imprecision in the recognition of emotions. This approach obtains an accuracy rate of 82%, suggesting a moderate degree of accuracy. The Attention-Based DenseNet (MABD) technique utilises the DenseNet architecture along with attention processes to enhance the recognition of essential features for emotion detection.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eInitially, applications of Convolutional Neural Networks (CNNs) for emotion detection, as demonstrated by Wen et al. in 2017, exhibited a comparatively low accuracy of 45.07%. This result indicates the early stage of CNN utilisation in this field during that period. On the other hand, a more up-to-date CNN implementation conducted by Shen and Xu in 2023 exhibits an essential enhancement in accuracy, reaching 93.33%. This highlights the progress made in CNN structures and training methods. Appasaheb Borgalli and Surve developed a specialised CNN architecture in 2022 that delivers a high accuracy rate of 91.58% for emotion recognition, geared to meet specific criteria. The Deep Neural Networks (DNNs) employed by An et al. in 2023 exhibit exceptional accuracy, reaching 96.97%. This result highlights their efficacy in collecting complex patterns within emotional data. The CNN approach proposed by Wen et al. in 2017, which incorporates enhancements, demonstrates minor improvements, achieving an accuracy rate of 50.7%.\u003c/p\u003e\n\u003cp\u003eFuzzy Classification methods, exemplified by the approach developed by Nicolai and Choi in 2016, integrate fuzzy logic with classification techniques to achieve an excellent accuracy level of 78.33%. In 2019, Kim et al. developed a Hierarchical Deep Neural Network Structure that effectively improves emotion recognition by utilising a hierarchical structure. This model achieved an impressive accuracy of 91.27%.\u003c/p\u003e\n\u003cp\u003eEnhanced CNN methodologies demonstrate exceptional performance. Zaman et al. achieved an accuracy of 98.01% in 2023, whereas Ramis et al. reached 98.29% in 2022. Similarly, in 2019, Wang et al. achieved an accuracy of 95.95%. In 2023, Benisha and Mirnalinee achieved an accuracy of 96.91%, while Podder et al. achieved 96.83%. These results demonstrate the exceptional performance of advanced CNN algorithms. In 2023, Chandra et al. developed a modified version of Attention-Based DenseNet (MABD), which demonstrated a lower accuracy of 42.99%. This indicates the presence of potential difficulties or areas that require additional improvement. The PSA-YOLO model, created by Zhang and Ma in 2023, employs the YOLO architecture with specific alterations to identify emotions. It has achieved an accuracy rate of 83.84%. The Stacked Sparse Auto-Encoder developed by Ahmad et al. in 2023 demonstrates a 92.5% success rate, indicating its efficacy in encoding and classifying emotions. The utilisation of pre-trained deep convolutional neural networks (DCNN) in transfer learning, as demonstrated by Akhand et al. in 2021, yields the maximum accuracy of 99.52% in the provided table, effectively showcasing the efficacy of pre-existing models.\u003c/p\u003e\n\u003cp\u003eSupervised learning approaches exhibit exceptional performance. In 2010, Chatterjee and Shi developed the Adaptive Neuro Fuzzy Inference System, which effectively combines neural networks and fuzzy logic to achieve a remarkable accuracy rate of 95%. The DLSANet, developed by Guo et al. in 2023, demonstrates an impressive accuracy rate of 93.81%. The HOG+ Random Forest algorithm, developed by Saeed in 2023, attains an accuracy of 92.97%.\u003c/p\u003e\n\u003cp\u003eK-Nearest Neighbours (KNN) approaches exhibit different degrees of accuracy. Shelke et al. achieved an accuracy of 77.29% in 2021, whereas Cornejo et al. achieved a far higher accuracy of 98.21% in 2015. In a study conducted by Harakannanavar et al. in 2023, they were able to attain an accuracy rate of 96.51%. In 2023, Subudhiray et al. demonstrated that combining KNN with Gabor filters, HOG features, and Local Binary Patterns (LBP) resulted in accuracies of 94.24%, 84.53%, and 90.65%, respectively.\u003c/p\u003e\n\u003cp\u003eThe multi-SVM algorithm, developed by Zhu et al. in 2015, exhibits a lesser accuracy rate of 61.97%. In contrast, the Softmax algorithm, proposed by Liu et al. in 2017, obtains a higher accuracy of 90%. Support Vector Machine (SVM) methods exhibit a variety of accuracy levels. In 2021, Jeen Retna Kumar\u0026nbsp;et al. achieved an accuracy of 97.3% by integrating Support Vector Machines (SVM) with cross-dataset analysis and feature improvement. In 2013, Zhou et al. attained an accuracy of 92.14%, while in 2017, Liu et al. achieved an accuracy of 89.6%. Lu et al. (2019) had a success rate of 98%, Cornejo et al. (2015) achieved a success rate of 82.86%, and Harakannanavar et al. (2023) achieved a success rate of 98.26%. 2019 Eng et al. demonstrated a reduced accuracy rate of 76.19%.\u003c/p\u003e\n\u003cp\u003eThe study that integrates Support Vector Machines (SVM) with Wavelet Coefficients (WC) produces a high accuracy rate of 94%. In contrast, the SVM Stacking method proposed by Delazeri et al. in 2022 only achieves an accuracy of 64.59%. Frank\u0026apos;s enhanced Support Vector Machine (SVM) technique shows a remarkable accuracy of 94.13%. The unsupervised learning methods developed by Tuncer et al. in 2023, such as PCA + LDA, show a remarkable accuracy of 97.09%. These techniques effectively combine dimensionality reduction and feature extraction.\u003c/p\u003e\n\u003cp\u003eThe table provides an overview of the notable progress made in emotion detection technology, demonstrating various approaches such as fuzzy inference systems, advanced neural network layouts, and conventional supervised learning methods. Over time, the precision of these methods has significantly enhanced, with modern techniques such as Transfer Learning + pre-trained DCNN and different Improved CNN models obtaining nearly perfect accuracy. This history demonstrates algorithms\u0026apos; continuous advancement and growing capacity to accurately represent and categorise human emotions.\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;2. Emotion detection by previous studies\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"605\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMethod\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eYear\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRemarks\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCategories\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eFuzzy emotion inference system\u0026nbsp;(31)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2019\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e6 emotions (Happy, anger, sad, disgust, fear and surprise)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eFuzzy Inference Engine\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eCNN\u0026nbsp;(32)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2017\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e45.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eCNN\u0026nbsp;(33)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e93.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eCustom CNN Architecture\u0026nbsp;(34)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2022\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e91.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eDNN\u0026nbsp;(35)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e96.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e4 emotions (angry, happy, neutral, fear)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eECNN\u0026nbsp;(32)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2017\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e50.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eFuzzy Classification\u0026nbsp;(36)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2016\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e78.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e6 emotions (Happy, anger, sad, disgust, fear, and surprise)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eHierarchical Deep Neural Network Structure\u0026nbsp;(37)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2019\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e91.27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e6 emotions (happy, anger, sad, disgust, fear, and surprise)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eImproved CNN\u0026nbsp;(38)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e98.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eImproved CNN\u0026nbsp;(39)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2022\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e98.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eImproved CNN\u0026nbsp;(40)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2019\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e95.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eImproved CNN\u0026nbsp;(41)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e96.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eImproved CNN\u0026nbsp;(42)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e96.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eModified Attention-Based DenseNet (MABD)\u0026nbsp;(43)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e42.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003ePSA-YOLO\u0026nbsp;(44)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e83.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eStacked Sparse Auto-Encoder\u0026nbsp;(45)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e92.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eTransfer learning + pre-trained DCNN\u0026nbsp;(46)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2021\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e99.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eNeural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eAdaptive Neuro Fuzzy Inference System\u0026nbsp;(47)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2010\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e5 emotions (happy, anger, sad, disgust and surprise)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eDLSANet\u0026nbsp;(48)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e93.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (Happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eHOG+ Random Forest\u0026nbsp;(49)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e92.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eKNN\u0026nbsp;(50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2021\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e77.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eKNN\u0026nbsp;(51)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2015\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e98.21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (Happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eKNN\u0026nbsp;(52)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e96.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e6 emotions (Happy, anger, sad, disgust, fear, and surprise)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eKNN+Gabor\u0026nbsp;(53)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e94.24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e6 emotions (happy, anger, sad, disgust, fear, and surprise)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eKNN+HOG\u0026nbsp;(53)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e84.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e6 emotions (happy, anger, sad, disgust, fear, and surprise)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eKNN+LBP\u0026nbsp;(53)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e90.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e6 emotions (happy, anger, sad, disgust, fear, and surprise)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eMulti-SVM\u0026nbsp;(54)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2015\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e61.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eSoftmax\u0026nbsp;(55)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2017\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eSVM\u0026nbsp;(56)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2021\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e97.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eSVM\u0026nbsp;(57)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2013\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e92.14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eSVM\u0026nbsp;(55)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2017\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e89.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eSVM\u0026nbsp;(50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2021\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eSVM\u0026nbsp;(58)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2019\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e6 emotions (Happy, anger, sad, disgust, fear, and surprise)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eSVM\u0026nbsp;(51)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2015\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e82.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (Happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eSVM\u0026nbsp;(52)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e98.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e6 emotions (Happy, anger, sad, disgust, fear, and surprise)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eSVM\u0026nbsp;(59)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2019\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e76.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (Happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003eSVM Stacking\u0026nbsp;(60)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2022\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e64.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (joy, anger, sad, disgust, fear, surprise and contempt)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eSupervised learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28.1457%;\"\u003e\n \u003cp\u003ePCA + LDA\u0026nbsp;(61)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.78146%;\"\u003e\n \u003cp\u003e2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.43709%;\"\u003e\n \u003cp\u003e97.09\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 35.9272%;\"\u003e\n \u003cp\u003e7 emotions (Happy, anger, sad, disgust, fear, surprise and neutral)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 18.7086%;\"\u003e\n \u003cp\u003eUnsupervised Learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;Researchers must address several current issues to improve the accuracy and reliability of facial emotion detection. The first is dependent setting. The low accuracy of emotion detection using facial expressions is caused by several issues, such as the dependent setting of expression, the distance of the face from the camera, obstacles in front of the face that cause face occlusion, and the low resolution of an image. Facial occlusion (62\u0026ndash;64), such as wearing sunglasses, a scarf or even a mask, is one critical factor that affects facial recognition performance (65,66). Moreover, according to Pham et al., distinguishing between emotions with similar facial expressions, such as fear and surprise, is complicated and causes low detection. However, a real-world scenario shows that dynamic environmental factors like facial obstacles, facial pose, and face and camera distance differences also cause low detection accuracy (20,26,65). Thus, as the main issue is low accuracy, tackling the issue of a dynamic environment with a general solution might help improve accuracy.\u003c/p\u003e"},{"header":"Material and Method ","content":"\u003cp\u003eThis section discusses the two types of datasets used and provides a brief explanation of our facial emotion detection framework.\u003c/p\u003e\n\u003cp\u003eMaterials\u003c/p\u003e\n\u003cp\u003eWe use two types of datasets in this study. The first dataset is the JAFFE dataset, a device-independent setting of facial expression images. The images were captured in a controlled environment, and the subjects were asked to mimic the expression. The second self-acquired dataset is Facial Expression Recognition 2013 (FER2013). This type of expression supports the real-world situation where no device setting can capture the expression. The dataset sample is shown in\u0026nbsp;Figure 1.\u003c/p\u003e\n\u003cp\u003eThe JAFFE image set\u0026apos;s first dataset included 213 images of ten Japanese female frontal facial photos, categorised into seven primary emotions (angry, disgust, fear, happy, neutral, sad and surprise) reflecting the system\u0026apos;s classification\u0026nbsp;(67,68). Each image in the JAFFE database has a resolution of 256x256 pixels. The images were taken under controlled conditions to guarantee consistency and reliability in the dataset. The JAFFEE database was designed and created by Michael Lyons, Miyuki Kamachi, and Jiro Gyoba\u0026nbsp;(22,69,70). The JAFFE image set\u0026apos;s first dataset included 213 images of ten Japanese female frontal facial photos, categorised into seven main emotions (angry, disgust, fear, happy, neutral, sad and surprise) reflecting the system\u0026apos;s classification\u0026nbsp;(67,68). Each image in the JAFFE database has a resolution of 256x256 pixels. The images were taken under controlled conditions to guarantee consistency and reliability in the dataset. The JAFFEE database was designed and created by Michael Lyons, Miyuki Kamachi, and Jiro Gyoba\u0026nbsp;(22,69,70).\u003c/p\u003e\n\u003cp\u003eWhile the second dataset is the FER2013 Dataset, as the study is focused on device-dependent setting facial expressions to imitate the real-world scenario, the latest version of Facial Expression Recognition 2013 (FER-2013) is used as facial data input. This expression exemplifies a scenario in the real world when there is no setup on any device that can capture that expression. The facial expression transition is more realistic, resembling the mimic expression captured in the laboratory (16). The dataset comprises grayscale images of faces with dimensions of 48x48 pixels. The facial features have been automatically aligned to ensure that the face is approximately centred and occupies an equivalent amount of space in each image. The objective is to classify each face according to the emotion shown in the facial expression into one of seven categories (0=Angry, 1=Disgust, 2=Fear, 3=Happy, 4=Sad, 5=Surprise, 6=Neutral). The training set has 28,709 samples, whereas the public test set comprises 3,589 examples.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eMethod Overview\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe block diagram of the proposed framework is shown in\u0026nbsp;Figure 2. The flowchart illustrates the process of facial emotion detection using two datasets: JAFFE and FER2013. The JAFFE dataset comprises facial images of Japanese female subjects displaying seven emotions, while the FER2013 dataset includes diverse subjects captured under various conditions, reflecting common issues in real-world images.\u003c/p\u003e\n\u003cp\u003eThe process starts with the JAFFE dataset serving as the training dataset. Feature extraction uses Wavelet Coefficient (WC) to extract meaningful features from each image. Wavelet Coefficients are beneficial as they effectively capture spatial and frequency information in the images, providing a robust representation of facial features crucial for accurate emotion detection. WC is good at resisting noises in images as no pre-processing is done. Moreover, WCs can reflect changes in grayscale images\u0026nbsp;(71), and to solve the illumination issues, WCs are less sensitive to variation in lighting, which helps ensure significant features are extracted\u0026nbsp;(72).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; The extracted features are then fed into a Support Vector Machine (SVM) to classify the images into seven emotions: Angry, Happy, Disgust, Sad, Fear, Surprise, and Neutral. SVM is one of the most used classifiers and can achieve high accuracy (71). SVM does perform well in terms of dimensionality and memory management (73). The extracted features are then fed into a Support Vector Machine (SVM) to classify the images into seven emotions: Angry, Happy, Disgust, Sad, Fear, Surprise, and Neutral. SVM is one of the most commonly used classifiers and can achieve high accuracy (71). SVM performs well in terms of dimensionality and memory management (73).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; The framework is subsequently tested using the FER2013 dataset, miming real-world situations with diverse subjects and varying conditions. This step assesses the effectiveness of the trained model in practical applications. The results and implications are discussed in the next section, evaluating the model\u0026apos;s performance in a more challenging and realistic setting.\u003c/p\u003e"},{"header":"Result and Discussion","content":"\u003cp\u003eThis section provides a comprehensive analysis of the results obtained from the experiments performed throughout the testing phase of the facial emotion detection framework.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eComparison of histogram intensities for seven emotions of JAFFE.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eImage histograms are frequently used throughout various image processing applications, including image enhancement, segmentation, and analysis\u0026nbsp;(27\u0026ndash;29). Histograms are graphic representations that illustrate the frequency of pixel occurrences for each intensity value in a picture, thereby displaying the distribution of pixel intensities. An 8-bit grayscale image has 256 distinct intensities, ranging from 0 to 255. These values facilitate an in-depth understanding of the image\u0026apos;s features, wherein several peaks may suggest the presence of distinct objects or regions in the image or offer insights into the general illumination, differences, and range of the image. The graph displays intensity values on the x-axis and the frequency or number of pixels at each intensity value on the y-axis\u0026nbsp;(30). Additionally, it can perform advanced analysis such as histogram equalisation, histogram specification, and local enhancement.\u0026nbsp;Figure 3\u0026nbsp;shows an example of histogram intensity analysis done towards the JAFFE dataset.\u003c/p\u003e\n\u003cp\u003eFor the Angry emotion, all three graphs exhibit a significant peak in the region of pixel value 50, indicating a high frequency of dark pixels. These obscure particles could indicate occluded areas or concealed facial features. At mid-range intensity, all three graphs exhibit several peaks. The frequency of pixel occurrences in the third graph is the lowest compared to the other two. Upon comparing the images, the facial expression displays lower signs of anger and less tension in the lips. These variations in facial features may result in distinct shadows, therefore implying a different underlying reason for the shape of the graph. On the other hand, the right side of the graph displays a uniform distribution of bright pixels throughout all three graphs.\u003c/p\u003e\n\u003cp\u003eMeanwhile, the dark pixel region across the disgust emotion analysis exhibits three distinct frequency occurrences. The first graph displays the highest number of peaks in this pattern, indicating a correlation with obscured facial features or shadowed areas. This suggests that the first set of images portrays more emotions of disgust compared to the other two sets of images. For mid-range intensities, the graph distribution was similar. However, there was a noticeable variation in the peak observed for image number one, indicating that this image displayed the highest level of severe disgust emotion compared to the other two images.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFinally, the bright pixels persist but are reduced in all three graphs. Meanwhile, the histogram graph shapes of the three images for the Fear emotion at mid-range intensities differ. In the first graph, the histogram displays a trough where fewer pixels are found within this range, indicating a lower degree of variability in the features. The second image demonstrates a more uniform distribution of pixels. This suggests that the characteristics within this range differ, indicating a less intense expression. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRegarding the third image, there is a constant distribution of mid-range intensities, indicating a balanced representation of facial features throughout this range. Meanwhile, the distribution of graph values for dark and bright pixels in these three graphs is subsequently the same. In the context of low intensities, the histogram exhibits a peak around a pixel value of 50, while for high intensities, the pixels are still present but less obvious.\u003c/p\u003e\n\u003cp\u003eThree images portraying the emotion of happy from the JAFFE dataset exhibit distinct styles of expression. The first and third images show the subject smiling with visible teeth. However, the second image does not display this characteristic. Regarding mid-range intensities, minor variations exist. The initial graph displays a histogram with a valley, indicating a reduced feature variation and a smaller number of pixels falling within this range. The second and third graphs\u0026nbsp;exhibit a more uniform and evenly distributed range of mid-level intensities. The pixel intensities of all three images demonstrate a consistent pattern, with the histogram displaying a flattened peak and suppressed pixel values starting from 200 onwards.\u003c/p\u003e\n\u003cp\u003eEach graph displays a single significant peak with a pixel value of roughly 50 for the dark intensities of the Neutral emotion histogram. All graphs display evenly distributed pixels, indicating a neutral expression. The histograms for mid-range intensities displayed uniform distribution across all graphs, suggesting a more subdued expression like neutrality. Finally, the bright intensities in all graphs are still visible but toned down, indicating neutrality. These illuminated areas suggest a neutral image.\u003c/p\u003e\n\u003cp\u003eThe dark pixels in all three sad images peak at a pixel value of 50 on the left side of the graphs. For mid-range intensities in the first graph, the middle section of the histogram displays a valley where fewer pixels are present, indicating reduced feature variability. Image two displays a more uniform distribution of pixel intensities, while the last image also shows an identical distribution, indicating balanced facial features within this range. For bright intensities around 200, the pixel intensities are visible but not as noticeable in all graphs.\u003c/p\u003e\n\u003cp\u003eFinally, the histograms of surprise expressions each peak at approximately pixel value 50, indicating low intensities of dark pixels. The dark\u0026nbsp;pixels may represent obscured features of the face or shaded areas caused by wrinkles on the forehead during a surprised expression. There are fewer pixels in this range for mid-range intensities in image number one, indicating less variability in features compared to images number\u0026nbsp;two and number\u0026nbsp;three. The mid-range intensities exhibit a uniform distribution. For bright pixels in all images, the pixels are evenly distributed, suggesting a predictable emotional state.\u003c/p\u003e\n\u003cp\u003eDistinct patterns are revealed by a closer look at histograms representing emotions like anger, disgust, fear, happy, neutrality, sad, and surprise. The patterns, with peaks and valleys in the histograms, represent changes in facial expressions and emotional intensity. Anger usually peaks at low intensities, but happy tends to have more consistent distributions. Furthermore, histograms provide vital insights into the intricate features of many emotional states, aiding in advancing emotion detection and understanding. However, this study found that angry emotion displays the most significant changes in pixel intensities, as shown by dark and mid-range regions compared to other emotions. The analysis of histograms provides an in-depth understanding of facial expressions and associated emotions. Studying the distribution of pixel intensities allows us to uncover essential details about human emotions, leading to progress in emotion detection and understanding.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo gain a deeper understanding of how different facial regions can give different meanings of expression for various emotions, the facial images were segmented into many areas of interest (ROI). The eyebrows, eyes, and mouth were selected as significant ROI. The pixel intensities are shown in\u0026nbsp;Figure 4.\u0026nbsp;From this experiment, the distribution of pixel intensities according to the selected ROI has unique and significant meanings across different emotions. More dark pixels in the eyebrow region may indicate an angry or sad emotion. The following experiment will confirm the detailed analysis.\u003c/p\u003e\n\u003cp\u003eAs previously discussed, angry emotions show significant changes in pixel intensities in the dark and middle-range regions. Thus, the image segmentation into three areas of interest revealed detailed differences for each emotion. The study found that the angry emotion had a high frequency of dark pixels, indicating significant shading due to furrowed eyebrows compared to other emotions. disgust, fear, and Happy also exhibit high intensities in the dark region, suggesting that these emotions are associated with eyebrow movements, such as furrowing or raising. In contrast with the surprise emotion, the graph skewed to the left, indicating an eyebrow movement that results in darker pixels.\u003c/p\u003e\n\u003cp\u003eSad and surprise are emotions typically closely associated with the eyes. When we compared the histogram of pixel intensities for all emotions, two distinct types of graphs emerged. First is the graph, which has a concentrated shading of dark pixels, stable mid-range intensities, and even a distribution of bright pixels, which brings angry, disgust, happy, sad, and neutral emotions. Fear and neutral show another graph pattern with less variable mid-range intensities, indicating consistent shading and evenly distributed bright pixels. This suggests that the eyes show changes only with specific emotions, suggesting that the eyes are more significant, smaller, or have reflection due to the tears.\u003c/p\u003e\n\u003cp\u003eFinally, the third ROI shows something different, with graphs showing three types of shapes. The first graph shows a left-skewed pattern, signifying a highly concentrated dark region. The second graph displays a consistent pixel distribution across all three sections (dark, middle and lighter pixels), and the final graph displays a right-skewed pattern. Given the visual significance of the mouth movement associated with surprise, this study shows that the skewed right graph corresponds to the emotion of surprise. The skewed right graph indicates a slight movement of the mouth in line with the emotions of anger, fear, and neutrality. However, disgust, Happy, and sad exhibit an equal distribution of pixels across all types of shading, indicating more significant mouth movement.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eEmotion Classification Using SVM on the JAFFE Dataset\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThis section describes the experimental results of SVM performances on the JAFFE dataset. Emotion recognition system sensitivity is the ability to precisely identify specific emotions. Following this measurement, the system is evaluated for its ability to accurately identify actual positive cases, meaning situations where the system reliably determines the existence of a particular emotion. It assesses how correctly the algorithm captures the actual instances of the emotion. High sensitivity in this context refers to the ability of an emotion detection system to accurately identify true expressions of Happy, sad, anger, and other emotions. Conversely, poor sensitivity suggests that the system may overlook certain instances of these emotions.\u003c/p\u003e\n\u003cp\u003eSpecificity is the ability of an emotion detection system to accurately identify persons who do not display a particular emotion. The system is evaluated on its capacity to identify true negative cases, namely when it accurately detects the lack of emotion. It measures the system\u0026apos;s ability to accurately classify occurrences that are not emotional, thus avoiding false positives. A highly specialised emotion detection system can accurately identify when a person is not displaying anger, sad, or any other specified emotion. On the other hand, a low specificity rate suggests that the system might inaccurately categorise non-emotional occurrences as the correct emotion.\u003c/p\u003e\n\u003cp\u003eTable 2\u0026nbsp;shows previous researchers\u0026apos; misclassification rate on JAFEE compared to our framework. Abbreviations represent the emotions: AN (Anger), DI (Disgust), FE (Fear), HA (Happy), NE (Neutral), SA (Sad), and SU. (Surprise). YOLO method shows higher misclassification rates overall, with the highest for surprise (39.0%) and the lowest for Happy (11.0%). Meanwhile, Stacked Sparse Auto-Encoder shows the lowest misclassification rates, with a good percentage of misclassification rate for most emotions, particularly surprise (1.0%) and Fear (1.0%). Our method shows a significantly higher misclassification rate for Happy, which is 33.0% and the lowest rate for Neutral emotion (1.0%). This is supported by the lowest result of specificity of Happy emotion from\u0026nbsp;Table 3, 61.82% and the lowest result of accuracy, 67.19%.\u003c/p\u003e\n\u003cp\u003eWhen sensitivity is low, causing the system to fail to identify a significant number of true positive emotions. As a result, individuals who genuinely express an emotion are missed by the emotion detection model. High false negatives occur when true positive instances are not correctly detected.\u003c/p\u003e\n\u003cp\u003eThe accuracy rates show how well an algorithm or system categorises different emotions. The method used accurately detects emotions of anger 92.19% of the time. The accuracy rate climbs to 96.88% when recognising disgust-related expressions, demonstrating an excellent capability in this domain. The model effectively manages the emotion of fear, reaching an accuracy rate of 93.75%. The accuracy decreases significantly to 67.19% while dealing with happy emotions. The model performs exceptionally well at identifying neutral emotions, achieving an accuracy rate of 98.44%. Sad, anger, and fear are categorised with an accuracy of 92.19%. Surprise emotions are identified with 93.75% accuracy. The accuracy rates offer essential insights into the model\u0026apos;s performance across multiple emotional categories. Achieving consistently high accuracy across all emotions can be challenging due to the complexity and subjectivity associated with human emotions.\u003c/p\u003e\n\u003cp\u003eThe specificity rates for different emotions offer insights into the model\u0026apos;s accuracy in distinguishing between various emotional states. The specificity rate for classifying anger is 100%, demonstrating a high level of accuracy in detecting anger, disgust, fear, neutrality, sad, and surprise without misclassifying other emotions. The specificity rate for happy is 61.82%, indicating that the proposed method has incorrectly categorised other emotions as Happy as the specificity rates provide essential information about the method\u0026apos;s capacity to minimise false positives and correctly categorise various emotional states. The rates are impacted by things including training data and model architecture.\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;3. Misclassification rate on JAFFE by previous researchers\u003c/p\u003e\n\u003cdiv align=\"Left\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"529\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"7\" style=\"width: 302px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMisclassification rate (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 227px;\"\u003e\n \u003cp\u003eClassification\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 43px;\"\u003e\n \u003cp\u003eAN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 43px;\"\u003e\n \u003cp\u003eDI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 43px;\"\u003e\n \u003cp\u003eFE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 43px;\"\u003e\n \u003cp\u003eHA.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 43px;\"\u003e\n \u003cp\u003eNE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 43px;\"\u003e\n \u003cp\u003eSA.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 43px;\"\u003e\n \u003cp\u003eSU.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 42.9924%;\"\u003e\n \u003cp\u003eYolo\u0026nbsp;(44)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e19.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e12.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e15.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e11.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e13.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e19.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e39.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 42.9924%;\"\u003e\n \u003cp\u003eCNN\u0026nbsp;(41)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e2.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e6.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e8.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e4.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e5.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e2.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e4.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 42.9924%;\"\u003e\n \u003cp\u003eStacked Sparse Auto-Encoder\u0026nbsp;(45)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e2.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e3.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e1.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e1.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e5.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e3.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e1.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 42.9924%;\"\u003e\n \u003cp\u003eSVM (This study)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e8.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e4.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e7.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e33.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e2.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e8.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8.14394%;\"\u003e\n \u003cp\u003e7.0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eTable\u0026nbsp;4. SVM classification performance measurement on JAFFE\u003c/p\u003e\n\u003cdiv align=\"Left\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 97px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eEmotion\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSensitivity (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 116px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpecificity (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 21.7978%;\"\u003e\n \u003cp\u003eAngry\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e92.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e37.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 21.7978%;\"\u003e\n \u003cp\u003eDisgust\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e96.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e75.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 21.7978%;\"\u003e\n \u003cp\u003eFear\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e93.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e63.64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 21.7978%;\"\u003e\n \u003cp\u003eHappy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e67.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e100.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e61.82\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 21.7978%;\"\u003e\n \u003cp\u003eNeutral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e98.44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e85.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 21.7978%;\"\u003e\n \u003cp\u003eSad\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e92.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e64.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 21.7978%;\"\u003e\n \u003cp\u003eSurprise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e93.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e42.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 26.0674%;\"\u003e\n \u003cp\u003e100\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u0026nbsp;\u003cem\u003eAnalysis of emotion detection on JAFFE\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThis section discussed performance results on the JAFFE dataset by comparing two methods. The first method uses only Support Vector Machines (SVM) in the framework, and the second method uses an additional Wavelet Coefficient together with SVM (WC+SVM). This experiment aims to see whether WC can enhance the classification performance for each emotion or not.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;5. Comparison of accuracy according to emotions\u003c/p\u003e\n\u003cdiv align=\"Left\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 69px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eEmotion\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 183px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy of SVM (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 202px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy of WC+SVM (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.1982%;\"\u003e\n \u003cp\u003eAngry\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40.3084%;\"\u003e\n \u003cp\u003e92.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 44.4934%;\"\u003e\n \u003cp\u003e96.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.1982%;\"\u003e\n \u003cp\u003eDisgust\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40.3084%;\"\u003e\n \u003cp\u003e96.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 44.4934%;\"\u003e\n \u003cp\u003e95.31\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.1982%;\"\u003e\n \u003cp\u003eFear\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40.3084%;\"\u003e\n \u003cp\u003e93.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 44.4934%;\"\u003e\n \u003cp\u003e95.31\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.1982%;\"\u003e\n \u003cp\u003eHappy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40.3084%;\"\u003e\n \u003cp\u003e67.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 44.4934%;\"\u003e\n \u003cp\u003e95.31\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.1982%;\"\u003e\n \u003cp\u003eNeutral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40.3084%;\"\u003e\n \u003cp\u003e98.44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 44.4934%;\"\u003e\n \u003cp\u003e96.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.1982%;\"\u003e\n \u003cp\u003eSad\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40.3084%;\"\u003e\n \u003cp\u003e92.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 44.4934%;\"\u003e\n \u003cp\u003e84.38\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.1982%;\"\u003e\n \u003cp\u003eSurprise\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40.3084%;\"\u003e\n \u003cp\u003e93.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 44.4934%;\"\u003e\n \u003cp\u003e95.31\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.1982%;\"\u003e\n \u003cp\u003eAverage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40.3084%;\"\u003e\n \u003cp\u003e90.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 44.4934%;\"\u003e\n \u003cp\u003e94.20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eSignificant differences in performance are observed when comparing SVM with WC+SVM across different emotional classes. WC+SVM yields a 96.88% accuracy rate in anger emotion, showing an enhancement compared to 92.19% accuracy using SVM. WC+SVM maintains a consistent accuracy rate of 95.31% for disgust and fear emotion, while the accuracy rate for SVM varies slightly around 96.88% and 93.75% for the respective emotions. There is a significant increment in the accuracy of happy emotion when using WC+SVM, achieving 95.31% accuracy compared to 67.19% only using SVM. For the accuracy rate of classification for sad emotion, WC+SVM has an accuracy rate of 84.3ST8%, whereas SVM retains a better accuracy rate of 92.19%. WC+SVM achieves accuracy rates like SVM for neutral and surprise emotions, with minor variations but no significant differences. WC+SVM obtains an accuracy of 96.88% for neutral emotion and 95.31% for surprise, slightly surpassing SVM\u0026apos;s accuracy. WC+SVM outperforms SVM with an average accuracy of 94.20% compared to SVM average accuracy of 90.10%.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe results indicate that using wavelet coefficient as feature extraction improves emotion detection accuracy, especially in identifying happy emotions. However, it may have a little lower accuracy for sad emotions. The Wavelet Coefficient technique is excellent at capturing spatial and spectral information. This ability is beneficial for identifying significant features in facial expressions, allowing for a more detailed investigation. Furthermore, this method is very efficient for individuals dealing with large datasets and aiming to extract thorough features.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eCross-database evaluation performance\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003ePrevious researchers frequently use a variety of well-known, publicly accessible facial expression databases to report on the accuracy of the suggested methodologies\u0026nbsp;(74).\u0026nbsp;Previous studies have demonstrated successful detection performance on particular datasets under intra-dataset testing scenarios\u0026nbsp;(75). According to Wen et al., a trained classifier might perform well on a specific set of training samples, but it may not be as effective in real-world scenarios where the sample size is much larger. Furthermore, real-world data usually causes a decrease in accuracy\u0026nbsp;(39,76). This is due to the lab-dependent setting, such as the expression, lighting and posture of the subject by the public facial databases\u0026nbsp;(43).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn the deep learning age, large-scale, high-quality datasets are an especially critical prerequisite for facial expression recognition (FER); nevertheless, most FER datasets are quite modest. The cross-datasets strategy is a popular approach to solving this issue\u0026nbsp;(77).\u0026nbsp;Previous research publications fail to conduct cross-database evaluations\u0026nbsp;(32,74,78). Thus, it is important to have a generalisation process where the classifier can perform well on unseen data and be more practical and valuable.\u0026nbsp;Nevertheless, only a few research studies do cross-dataset testing evaluations to demonstrate the reliability of facial expressions identified by unfamiliar sources.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eHybrid approaches combining several descriptors, face regions, picture modalities, and visual cues, including head movement and body gestures, need more research to enhance the effectiveness of facial expression identification. In the deep learning age, large-scale, high-quality datasets are an especially critical prerequisite for facial expression recognition (FER); nevertheless, most FER datasets are quite modest. The cross-datasets strategy is a popular approach to solving this issue\u0026nbsp;(77).\u0026nbsp;Previous research publications fail to conduct cross-database evaluations\u0026nbsp;(32,74,78). Thus, it is important to have a generalisation process where the classifier can perform well on unseen data and be more practical and valuable.\u0026nbsp;Nevertheless, only a few research studies do cross-dataset testing evaluations to demonstrate the reliability of facial expressions identified by unfamiliar sources. Hybrid approaches combining several descriptors, face regions, picture modalities, and visual cues, including head movement and body gestures, need more research to enhance the effectiveness of facial expression identification.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eNowadays, most FER methods are developed and assessed with the presumption that the dataset containing the training and testing facial expression images will be the same\u0026nbsp;(76). A periodic facial expression recognition challenge might offer a reasonable basis for this comparison. It would provide a clearer picture of the advancement of the field and enable us to establish new objectives, obstacles, and targets.\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;5\u0026nbsp;compares the numbers of different methods for facial emotion detection, outlining different classification methods, different sets of training and testing datasets and average accuracy achieved. The table gives a general overview of the several techniques used for emotion intensity analysis, including ECNN, CNN, Gabor with hierarchical classifier, SVM, RF and SVM, Local Directional Pattern with SVM, Boosted-LBP, Transfer Learning with SVM, and Wavelet Coefficient with SVM. Most authors used seven common types of emotion except Mayer\u0026nbsp;(79),\u0026nbsp;where the research only used six basic emotions that excluded neutral emotion.\u003c/p\u003e\n\u003cp\u003eThere are noticeable differences in accuracy and performance between the given approaches for various datasets. For example, the ECNN approach obtained an accuracy of 76.05% when evaluated on CK+ and trained on the FER2013 dataset\u0026nbsp;(32). Comparably, the CNN approach achieved an accuracy of 73.38% after being tested on CK+ but trained on the same dataset\u0026nbsp;(32). Nevertheless, the accuracy of these techniques significantly dropped when they were used on the JAFFE dataset, indicating possible difficulties in generalising across various datasets.\u003c/p\u003e\n\u003cp\u003eIn contrast to the CNN and ECNN approaches, the Gabor with hierarchical classifier method, evaluated on CK+ and trained on the JAFFE dataset, they demonstrated an acceptable performance standard with an accuracy of 54.05%\u0026nbsp;Gu et al., 2012). Conversely, the accuracies of SVM-based techniques, such as those by Mayer et al., varied according to the training dataset used. For example, SVM obtained an accuracy of 66.20% when evaluated on CK+ and trained on the MMI dataset\u0026nbsp;(79).\u003c/p\u003e\n\u003cp\u003eAfter training on the BU-3DFE dataset and testing on JAFFE, an SVM-based method achieved an accuracy of 41.96%, suggesting lower accuracy on this dataset. Zhu et al. suggested Transfer Learning with SVM showed varying accuracies depending on whether the data was normalised\u0026nbsp;(54). The normalised strategy achieved an accuracy of 60.09% when tested on JAFFE and trained on the FEED dataset, while the non-normalised approach achieved an accuracy of 46.48%.\u003c/p\u003e\n\u003cp\u003eSpecifically, our suggested approach, Wavelet Coefficient with SVM, showed encouraging outcomes, averaging an extraordinary 78.29% accuracy. Our solution outperformed numerous other methods in the table, demonstrating robust performance after training on the JAFFE dataset and testing on an occluded subset of FER2013. This indicates the effectiveness of using SVM classification in conjunction with wavelet coefficient analysis for tasks related to emotion intensity analysis.\u003c/p\u003e\n\u003cp\u003eTable\u0026nbsp;6. Performance comparison between our framework and previous works that use cross-database method\u003c/p\u003e\n\u003cdiv align=\"Left\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"605\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 35.8678%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMethod\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAuthors\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTraining Dataset\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTesting Dataset\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAverage Accuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eWavelet Coefficient with SVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003eThis study\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eJAFFE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eOccluded FER2013\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e78.29\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eECNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(32)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eFER2013\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eCK+\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e76.05\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eCNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(32)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eFER2013\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eCK+\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e73.38\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eGabor with hierarchical classifier\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(80)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eJAFFE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eCK+\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e54.05\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(79)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eMMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eCK+\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e66.20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(79)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eFEED\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eCK+\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e56.60\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eECNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(32)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eFER2013\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eJAFFE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e50.70\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eCNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(32)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eFER2013\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eJAFFE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e45.07\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eGabor with hierarchical classifier\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(80)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eCK+\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eJAFFE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e55.87\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eRF and SVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(81)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eBU-3DFE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eJAFFE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e41.96\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eLocal Directional Pattern with SVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(57)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eCK\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eJAFFE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e45.71\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eBoosted-LBP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(82)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eCK\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eJAFFE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e41.30\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eTransfer Learning with SVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(54)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eFEED\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eJAFFE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e46.48\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 35.8678%;\"\u003e\n \u003cp\u003eTransfer Learning with SVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e(54)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eFEED\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 19.1736%;\"\u003e\n \u003cp\u003eJAFFE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.8926%;\"\u003e\n \u003cp\u003e60.09\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eOur research demonstrates how well our proposed method can identify different facial emotions. We trained our model using the JAFFE dataset. Then, we tested it on obscured data from the FER2013 dataset, an excellent representation of the data we\u0026rsquo;d encounter in everyday life, as it closely resembles real-world scenarios. The key takeaway from the table is the importance of considering factors like the methodology used and variations in the datasets when evaluating the performance of emotion intensity analysis methods. This study needs deeper research and development in this area. We need to understand that specific methods are only suitable for certain situations or issues; thus, there is a need to have a technique that can generally overcome all the problems simultaneously.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eOur research demonstrates that wavelet coefficient analysis with Support Vector Machine (SVM) classification provides a scalable and effective solution for detecting emotion intensities, particularly within large datasets. This framework\u0026rsquo;s robustness in handling occlusions makes it well-suited for real-time applications, such as video surveillance, where faces may be partially obscured, and real-time data processing is essential. To assess detection accuracy, we combined the wavelet coefficient with a cross-database method, testing its performance across various datasets, including the JAFFE dataset, which contains seven emotions, and FER2013, which closely mimics real-world conditions. We examined the impact of pixel density adjustments before applying wavelet analysis to optimise detection accuracy, showing that the framework consistently outperforms existing methods. However, it has yet to achieve 100% accuracy.\u003c/p\u003e \u003cp\u003eOur approach shows strong potential for applications in psychological studies and social media emotion analysis, where scalable data processing and adaptability are critical. Experimental results confirm the framework\u0026rsquo;s ability to navigate the intricacies of emotion detection across varied, occluded datasets, supporting its viability in big data-driven environments. To further refine the accuracy of emotion representation, we plan to integrate additional feature extraction and selection methods in future studies, along with testing on more wild, real-world databases. This will ensure that the framework remains robust across diverse data types and comprehensively addresses current facial emotion detection challenges.\u003c/p\u003e"},{"header":"Abbreviation","content":"\u003cp\u003eSupport Vector Machine (SVM), Wavelet Coefficient (WC), Japanese Female Facial Expression (JAFFE), Facial Expression Recognition 2013 (FER2013), AN (Anger), DI (Disgust), FE (Fear), HA (Happy), NE (Neutral), SA (Sad), and SU. (Surprise)\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthors' contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eASA designed and wrote the manuscript. RH contributed to the design and critically reviewed the manuscript. NHZ, SHM, ARK, MLMS, and SK participated in reviewing the manuscript to enhance its clarity and coherence. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Ministry of Higher Education, Malaysia (MOHE) under the Fundamental Research Grant Scheme (FRGS), FRGS/1/2023/ICT02/UTM/02/8.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eKumar A, Kaur A, Kumar M. Face detection techniques: a review. Artif Intell Rev [Internet]. 2019 Aug 15 [cited 2020 Nov 30];52(2):927\u0026ndash;48. Available from: https://doi.org/10.1007/s10462-018-9650-2\u003c/li\u003e\n \u003cli\u003eRevina IM, Emmanuel WRS. A Survey on Human Face Expression Recognition Techniques. J King Saud Univ - Comput Inf Sci [Internet]. 2018; Available from: https://doi.org/10.1016/j.jksuci.2018.09.002\u003c/li\u003e\n \u003cli\u003eCowen A, Sauter D, Tracy JL, Keltner D. Mapping the Passions: Toward a High-Dimensional Taxonomy of Emotional Experience and Expression. Psychol Sci Public Interes [Internet]. 2019 Jul 17 [cited 2019 Oct 15];20(1):69\u0026ndash;90. Available from: http://journals.sagepub.com/doi/10.1177/1529100619850176\u003c/li\u003e\n \u003cli\u003eEkman P. Facial expression and emotion. Am Psychol [Internet]. 1993 [cited 2020 Nov 25];48(4):384\u0026ndash;92. Available from: /record/1993-32252-001\u003c/li\u003e\n \u003cli\u003eEkman P. Facial expressions of emotion: an old controversy and new findings. [Internet]. Vol. 335, Philosophical transactions of the Royal Society of London. Series B, Biological sciences. Philos Trans R Soc Lond B Biol Sci; 1992 [cited 2020 Nov 25]. p. 63\u0026ndash;9. Available from: https://pubmed.ncbi.nlm.nih.gov/1348139/\u003c/li\u003e\n \u003cli\u003eEkman P, Freisen W V., Ancoli S. Facial signs of emotional experience. J Pers Soc Psychol [Internet]. 1980 Dec [cited 2020 Nov 25];39(6):1125\u0026ndash;34. Available from: /record/1981-25797-001\u003c/li\u003e\n \u003cli\u003eBahreini K, van der Vegt W, Westera W. A fuzzy logic approach to reliable real-time recognition of facial emotions. Multimed Tools Appl [Internet]. 2019 Jul 30 [cited 2020 Mar 6];78(14):18943\u0026ndash;66. Available from: https://doi.org/10.1007/s11042-019-7250-z\u003c/li\u003e\n \u003cli\u003eDolan E, Hancock E, Wareing A. An evaluation of online learning to teach practical competencies in undergraduate health science students. Internet High Educ. 2015 Jan 1;24:21\u0026ndash;5.\u003c/li\u003e\n \u003cli\u003eKumari J, Rajesh R, Pooja KM. Facial Expression Recognition: A Survey. Procedia Comput Sci [Internet]. 2015 Jan 1 [cited 2020 Dec 3];58:486\u0026ndash;91. Available from: http://dx.doi.org/10.1016/j.procs.2015.08.011\u003c/li\u003e\n \u003cli\u003ePantic M, Valstar M, Rademaker R, Maat L. Web-based database for facial expression analysis. In: IEEE International Conference on Multimedia and Expo, ICME 2005. 2005. p. 317\u0026ndash;21.\u003c/li\u003e\n \u003cli\u003eYin L, Wei X, Sun Y, Wang J, Rosato MJ. A 3D facial expression database for facial behavior research. In: FGR 2006: Proceedings of the 7th International Conference on Automatic Face and Gesture Recognition. 2006. p. 211\u0026ndash;6.\u003c/li\u003e\n \u003cli\u003eGross R, Matthews I, Cohn J, Kanade T, Baker S. Multi-PIE. Image Vis Comput. 2010 May 1;28(5):807\u0026ndash;13.\u003c/li\u003e\n \u003cli\u003eLucey P, Cohn JF, Kanade T, Saragih J, Ambadar Z, Matthews I. The extended Cohn-Kanade dataset (CK+): A complete dataset for action unit and emotion-specified expression. In: 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Workshops, CVPRW 2010. 2010. p. 94\u0026ndash;101.\u003c/li\u003e\n \u003cli\u003eLangner O, Dotsch R, Bijlstra G, Wigboldus DHJ, Hawk ST, van Knippenberg A. Presentation and validation of the radboud faces database. Cogn Emot [Internet]. 2010 Dec [cited 2020 Dec 3];24(8):1377\u0026ndash;88. Available from: https://www.tandfonline.com/doi/abs/10.1080/02699930903485076\u003c/li\u003e\n \u003cli\u003eValstar MF, Jiang B, Mehu M, Pantic M, Scherer K. The first facial expression recognition and analysis challenge. In: 2011 IEEE International Conference on Automatic Face and Gesture Recognition and Workshops, FG 2011. 2011. p. 921\u0026ndash;6.\u003c/li\u003e\n \u003cli\u003eDhall A, Goecke R, Lucey S, Gedeon T. Static facial expression analysis in tough conditions: Data, evaluation protocol and benchmark. In: Proceedings of the IEEE International Conference on Computer Vision. 2011. p. 2106\u0026ndash;12.\u003c/li\u003e\n \u003cli\u003eZhao G, Huang X, Taini M, Li SZ, Pietik\u0026auml;inen M. Facial expression recognition from near-infrared videos. Image Vis Comput. 2011 Aug 1;29(9):607\u0026ndash;19.\u003c/li\u003e\n \u003cli\u003eGoodfellow IJ, Erhan D, Carrier PL, Courville A, Mirza M, Hamner B, et al. Challenges in representation learning: A report on three machine learning contests. In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) [Internet]. Springer, Berlin, Heidelberg; 2013 [cited 2020 Dec 3]. p. 117\u0026ndash;24. Available from: http://www.kaggle.com/c/challenges-in-representation-learning-facial-\u003c/li\u003e\n \u003cli\u003eYan W-J, Li X, Wang S-J, Zhao G, Liu Y-J, Chen Y-H, et al. CASME II: An Improved Spontaneous Micro-Expression Database and the Baseline Evaluation. Guo K, editor. PLoS One [Internet]. 2014 Jan 27 [cited 2020 Dec 3];9(1):e86041. Available from: https://dx.plos.org/10.1371/journal.pone.0086041\u003c/li\u003e\n \u003cli\u003eLi S, Deng W, Du JP. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In: Proceedings - 30th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017. Institute of Electrical and Electronics Engineers Inc.; 2017. p. 2584\u0026ndash;93.\u003c/li\u003e\n \u003cli\u003eMollahosseini A, Hasani B, Mahoor MH. AffectNet: A Database for Facial Expression, Valence, and Arousal Computing in the Wild. IEEE Trans Affect Comput. 2019 Jan 1;10(1):18\u0026ndash;31.\u003c/li\u003e\n \u003cli\u003eLyons MJ, Kamachi M, Gyoba J. Coding Facial Expressions with Gabor Wavelets (IVC Special Issue). 2020 Sep 13 [cited 2020 Dec 3]; Available from: http://arxiv.org/abs/2009.05938\u003c/li\u003e\n \u003cli\u003ePoulose A, Kim JH, Han DS. Feature Vector Extraction Technique for Facial Emotion Recognition Using Facial Landmarks. Int Conf ICT Converg. 2021;2021-October:1072\u0026ndash;6.\u003c/li\u003e\n \u003cli\u003eKishan Kondaveeti H, Vishal Goud M. Emotion Detection using Deep Facial Features. Proc IEEE Int Conf Advent Trends Multidiscip Res Innov ICATMRI 2020. 2020 Dec 30;\u003c/li\u003e\n \u003cli\u003eSingh S, Prasad SVAV. Techniques and challenges of face recognition: A critical review. In: Procedia Computer Science. Elsevier BV; 2018. p. 536\u0026ndash;43.\u003c/li\u003e\n \u003cli\u003ePham TTD, Kim S, Lu Y, Jung SW, Won CS Facial Action Units-Based Image Retrieval for Facial Expression Recognition. IEEE Access. 2019;7:5200\u0026ndash;7.\u003c/li\u003e\n \u003cli\u003eTrongtirakul T, Phanthuna N. Image Enhancement Using Weighted Bi-Histogram Equalization. Int J Appl Math Informatics. 2021;15:98\u0026ndash;101.\u003c/li\u003e\n \u003cli\u003eBarik R, Saha T, Chatterjee S. Reduction based histogram equalisation technique for image enhancement. Int J Innov Technol Explor Eng [Internet]. 2019 [cited 2024 Feb 15];8(11 Special Issue):1268\u0026ndash;71. Available from: http://www.mathworks.com/products/demos/#\u003c/li\u003e\n \u003cli\u003eJumiawi WAH, El-Zaart A. Improvement in the Between-Class Variance Based on Lognormal Distribution for Accurate Image Segmentation. Entropy [Internet]. 2022 Aug 29 [cited 2024 Feb 15];24(9):1204. Available from: https://www.mdpi.com/1099-4300/24/9/1204/htm\u003c/li\u003e\n \u003cli\u003eSysko LR, Davis MA. From image to data using common image-processing techniques. Curr Protoc Cytom [Internet]. 2010 Oct 1 [cited 2024 Feb 15];54(SUPPL.54). Available from: https://typeset.io/papers/from-image-to-data-using-common-image-processing-techniques-2pwtgq1yha\u003c/li\u003e\n \u003cli\u003eLiliana DY, Basaruddin T, Widyanto MR, Oriza IID. Fuzzy emotion: a natural approach to automatic facial expression recognition from psychological perspective using fuzzy system. Cogn Process [Internet]. 2019 Nov 1 [cited 2020 Mar 6];20(4):391\u0026ndash;403. Available from: http://www.ncbi.nlm.nih.gov/pubmed/31209637\u003c/li\u003e\n \u003cli\u003eWen G, Hou Z, Li H, Li D, Jiang L, Xun E. Ensemble of Deep Neural Networks with Probability-Based Fusion for Facial Expression Recognition. Cognit Comput. 2017 Oct 1;9(5):597\u0026ndash;610.\u003c/li\u003e\n \u003cli\u003eShen T, Xu H. Facial Expression Recognition Based onMulti-Channel Attention Residual Network. C - Comput Model Eng Sci [Internet]. 2023 Sep 29 [cited 2024 Mar 14];135(1):539\u0026ndash;60. Available from: https://www.techscience.com/CMES/v135n1/50070/html\u003c/li\u003e\n \u003cli\u003eAppasaheb Borgalli MR, Surve D.S. Deep learning for facial emotion recognition using custom CNN architecture. In: Journal of Physics: Conference Series. IOP Publishing; 2022. p. 12004.\u003c/li\u003e\n \u003cli\u003eAn Y, Lee J, Bak ES, Pan S. Deep Facial Emotion Recognition Using Local Features Based on Facial Landmarks for Security System. Comput Mater Contin [Internet]. 2023 Aug 30 [cited 2023 Dec 8];76(2):1817\u0026ndash;32. Available from: https://www.techscience.com/cmc/v76n2/54000/html\u003c/li\u003e\n \u003cli\u003eNicolai A, Choi A. Facial Emotion Recognition Using Fuzzy Systems. In: Proceedings - 2015 IEEE International Conference on Systems, Man, and Cybernetics, SMC 2015. Institute of Electrical and Electronics Engineers Inc.; 2016. p. 2216\u0026ndash;21.\u003c/li\u003e\n \u003cli\u003eKim JH, Kim BG, Roy PP, Jeong DM. Efficient facial expression recognition algorithm based on hierarchical deep neural network structure. IEEE Access. 2019;7:41273\u0026ndash;85.\u003c/li\u003e\n \u003cli\u003eZaman K, Zhaoyun S, Shah B, Hussain T, Shah SM, Ali F, et al. A novel driver emotion recognition system based on deep ensemble classification. Complex Intell Syst [Internet]. 2023 [cited 2023 Dec 25];9(6):6927\u0026ndash;52. Available from: https://doi.org/10.1007/s40747-023-01100-9\u003c/li\u003e\n \u003cli\u003eRamis S, Buades JM, Perales FJ, Manresa-Yee C. A Novel Approach to Cross dataset studies in Facial Expression Recognition. Multimed Tools Appl [Internet]. 2022 Nov 1 [cited 2024 Feb 29];81(27):39507\u0026ndash;44. Available from: https://link.springer.com/article/10.1007/s11042-022-13117-2\u003c/li\u003e\n \u003cli\u003eWang Y, Li Y, Song Y, Rong X. Facial expression recognition based on auxiliary models. Algorithms [Internet]. 2019 Oct 31 [cited 2019 Dec 7];12(11):227. Available from: https://www.mdpi.com/1999-4893/12/11/227\u003c/li\u003e\n \u003cli\u003eBenisha S, Mirnalinee TT. Human Facial Emotion Recognition using Deep Neural Networks. Int Arab J Inf Technol [Internet]. 2023 [cited 2024 Mar 14];20(3):303\u0026ndash;9. Available from: https://doi.org/10.34028/iajit/20/3/2\u003c/li\u003e\n \u003cli\u003ePodder T, Bhattacharya D, Majumder P, Balas VE. A feature boosted deep learning method for automatic facial expression recognition. PeerJ Comput Sci [Internet]. 2023 Jan 31 [cited 2024 Mar 14];9:e1216. Available from: https://peerj.com/articles/cs-1216\u003c/li\u003e\n \u003cli\u003eChandra J, Annappa B, Rashmi Adyapady R. Cross-Database Facial Expression Recognition using CNN with Attention Mechanism. In: 2023 14th International Conference on Computing Communication and Networking Technologies, ICCCNT 2023. Institute of Electrical and Electronics Engineers Inc.; 2023.\u003c/li\u003e\n \u003cli\u003eZhang R, Ma R. Facial expression recognition method based on PSA\u0026mdash;YOLO network. Front Neurorobot. 2023 Jan 17;16:1057983.\u003c/li\u003e\n \u003cli\u003eAhmad M, Saira, Alfandi O, Khattak AM, Qadri SF, Saeed IA, et al. Facial expression recognition using lightweight deep learning modeling. Math Biosci Eng [Internet]. 2023 [cited 2024 Mar 14];20(5):8208\u0026ndash;25. Available from: http://www.aimspress.com/article/doi/10.3934/mbe.2023357\u003c/li\u003e\n \u003cli\u003eAkhand MAH, Roy S, Siddique N, Kamal MAS, Shimamura T. Facial emotion recognition using transfer learning in the deep CNN. Electron [Internet]. 2021 Apr 27 [cited 2022 Feb 16];10(9):1036. Available from: https://www.mdpi.com/2079-9292/10/9/1036/htm\u003c/li\u003e\n \u003cli\u003eChatterjee S, Shi H. A novel Neuro Fuzzy approach to human emotion determination. In: Proceedings - 2010 Digital Image Computing: Techniques and Applications, DICTA 2010. 2010. p. 282\u0026ndash;7.\u003c/li\u003e\n \u003cli\u003eGuo X, Lu S, Wang S, Lu Z, Zhang Y. DLSANet: Facial expression recognition with double-code LBP-layer spatial-attention network. IET Image Process [Internet]. 2023 Jul 1 [cited 2024 Mar 14];17(9):2659\u0026ndash;72. Available from: https://onlinelibrary.wiley.com/doi/full/10.1049/ipr2.12817\u003c/li\u003e\n \u003cli\u003eSaeed VA. A Framework for Recognition of Facial Expression Using HOG Features. Int J Math Stat Comput Sci [Internet]. 2023 May 27 [cited 2024 Mar 14];2:1\u0026ndash;8. Available from: https://ijmscs.org/index.php/ijmscs/article/view/7815\u003c/li\u003e\n \u003cli\u003eShelke N, Upadhye S, Uparkar SS, Bawane H, Shivvanshi D, Kosekar S. Approaches for harvesting on emotion extraction from human facial expressions. Indian J Comput Sci Eng. 2021 Jul 1;12(4):921\u0026ndash;44.\u003c/li\u003e\n \u003cli\u003eCornejo JYR, Pedrini H, Fl\u0026oacute;rez-Revuelta F. Facial expression recognition with occlusions based on geometric representation. In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) [Internet]. Springer Verlag; 2015 [cited 2024 Mar 14]. p. 263\u0026ndash;70. Available from: https://link.springer.com/chapter/10.1007/978-3-319-25751-8_32\u003c/li\u003e\n \u003cli\u003eHarakannanavar SS, Sapnakumari C, Ramachandra AC, Pramodhini R, Prashanth CR Performance Evaluation of Fusion Based Efficient Algorithm for Facial Expression Recognition. Indian J Sci Technol [Internet]. 2023 Jan 28 [cited 2024 Mar 14];16(4):266\u0026ndash;76. Available from: https://indjst.org/articles/performance-evaluation-of-fusion-based-efficient-algorithm-for-facial-expression-recognition\u003c/li\u003e\n \u003cli\u003eSubudhiray S, Palo HK, Das N. K-nearest neighbor based facial emotion recognition using effective features. IAES Int J Artif Intell [Internet]. 2023 Mar 1 [cited 2024 Mar 14];12(1):57. Available from: https://ijai.iaescore.com/index.php/IJAI/article/view/21156\u003c/li\u003e\n \u003cli\u003eZhu R, Zhang T, Zhao Q, Wu Z. A transfer learning approach to cross-database facial expression recognition. In: Proceedings of 2015 International Conference on Biometrics, ICB 2015. Institute of Electrical and Electronics Engineers Inc.; 2015. p. 293\u0026ndash;8.\u003c/li\u003e\n \u003cli\u003eLiu Y, Li Y, Ma X, Song R. Facial expression recognition with fusion features extracted from salient facial areas. Sensors (Switzerland). 2017 Apr 1;17(4).\u003c/li\u003e\n \u003cli\u003eJeen Retna Kumar R, Sundaram M, Arumugam N, Kavitha V. Face feature extraction for emotion recognition using statistical parameters from subband selective multilevel stationary biorthogonal wavelet transform. Soft Comput [Internet]. 2021 Apr 1 [cited 2022 Sep 7];25(7):5483\u0026ndash;501. Available from: https://dl.acm.org/doi/10.1007/s00500-020-05550-y\u003c/li\u003e\n \u003cli\u003eZhou J, Xu T, Gan J. Feature extraction based on local directional pattern with SVM decision-level fusion for facial expression recognition. Int J Bio-Science Bio-Technology. 2013;5(2):101\u0026ndash;10.\u003c/li\u003e\n \u003cli\u003eLu Y, Wang S, Zhao W. Facial expression recognition based on discrete separable shearlet transform and feature selection. Algorithms [Internet]. 2019 Dec 31 [cited 2019 Dec 8];12(1):11. Available from: http://www.mdpi.com/1999-4893/12/1/11\u003c/li\u003e\n \u003cli\u003eEng SK, Ali H, Cheah AY, Chong YF. Facial expression recognition in JAFFE and KDEF Datasets using histogram of oriented gradients and support vector machine. In: IOP Conference Series: Materials Science and Engineering [Internet]. IOP Publishing; 2019 [cited 2024 Mar 17]. p. 012031. Available from: https://iopscience.iop.org/article/10.1088/1757-899X/705/1/012031\u003c/li\u003e\n \u003cli\u003eDelazeri BR, Vera LL, Barddal JP, Koerich AL, Alceu De SB. Evaluation of Self-taught Learning-based Representations for Facial Emotion Recognition. In: Proceedings of the International Joint Conference on Neural Networks. 2022.\u003c/li\u003e\n \u003cli\u003eTuncer T, Dogan S, Subasi A. Automated facial expression recognition using novel textural transformation. J Ambient Intell Humaniz Comput [Internet]. 2023 Jul 1 [cited 2024 Mar 14];14(7):9435\u0026ndash;49. Available from: https://link.springer.com/article/10.1007/s12652-023-04612-x\u003c/li\u003e\n \u003cli\u003eYang B, Wu J, Ikeda K, Hattori G, Sugano M, Iwasawa Y, et al. Face-mask-aware Facial Expression Recognition based on Face Parsing and Vision Transformer. Pattern Recognit Lett [Internet]. 2022 Dec 1 [cited 2024 Jun 18];164:173\u0026ndash;82. Available from: https://dl.acm.org/doi/10.1016/j.patrec.2022.11.004\u003c/li\u003e\n \u003cli\u003eSaeed U. Facial micro-expressions as a soft biometric for person recognition. Pattern Recognit Lett [Internet]. 2021 Mar 1 [cited 2024 Jun 18];143:95\u0026ndash;103. Available from: https://dl.acm.org/doi/10.1016/j.patrec.2020.12.021\u003c/li\u003e\n \u003cli\u003eAhmad AS, Hassan R, Zakaria NH, Moi SH. Comparative studies of facial emotion detection in online learning. In: AIP Conference Proceedings [Internet]. American Institute of Physics Inc.; 2023 [cited 2024 Jan 24]. Available from: /aip/acp/article/2827/1/030009/2910737/Comparative-studies-of-facial-emotion-detection-in\u003c/li\u003e\n \u003cli\u003eCanedo D, Neves AJR. Facial expression recognition using computer vision: A systematic review [Internet]. Vol. 9, Applied Sciences (Switzerland). MDPI AG; 2019 [cited 2019 Dec 7]. p. 4678. Available from: www.mdpi.com/journal/applsci\u003c/li\u003e\n \u003cli\u003eSu Y, Yang Y, Guo Z, Yang W. Face recognition with occlusion. In: Proceedings - 3rd IAPR Asian Conference on Pattern Recognition, ACPR 2015. Institute of Electrical and Electronics Engineers Inc.; 2016. p. 670\u0026ndash;4.\u003c/li\u003e\n \u003cli\u003eRevina ImM, Emmanuel WRS, Sam Emmanuel W. A Survey on Human Face Expression Recognition Techniques. J King Saud Univ - Comput Inf Sci [Internet]. 2018 [cited 2023 Jun 14]; Available from: https://doi.org/10.1016/j.jksuci.2018.09.002\u003c/li\u003e\n \u003cli\u003eKasim S, Hassan R, Zaini NHNH, Ahmad ASAS, Ramli AAAA, Saedudin RRRR. A Study on Facial Expression Recognition Using Local Binary Pattern. Int J Adv Sci Eng Inf Technol [Internet]. 2017 Oct 26 [cited 2019 May 29];7(5):1621\u0026ndash;6. Available from: http://ijaseit.insightsociety.org/index.php?option=com_content\u0026amp;view=article\u0026amp;id=9\u0026amp;Itemid=1\u0026amp;article_id=3390\u003c/li\u003e\n \u003cli\u003ePutra WB, Arifin F. Real-Time Emotion Recognition System to Monitor Student\u0026rsquo;s Mood in a Classroom. In: Journal of Physics: Conference Series. 2019.\u003c/li\u003e\n \u003cli\u003eLyons M, Akamatsu S, Kamachi M, Gyoba J. Coding facial expressions with Gabor wavelets. In: Proceedings - 3rd IEEE International Conference on Automatic Face and Gesture Recognition, FG 1998. IEEE Computer Society; 1998. p. 200\u0026ndash;5.\u003c/li\u003e\n \u003cli\u003eLiu C, Wang A. Adaptive High-order Variation De-noising Method for Edge Detection with Wavelet Coefficients. KSII Trans Internet Inf Syst [Internet]. 2023 Feb 28 [cited 2024 Jun 13];17(2):412\u0026ndash;34. Available from: https://link.springer.com/article/10.1007/s11432-006-0494-z\u003c/li\u003e\n \u003cli\u003eHu H. Illumination invariant face recognition based on dual-tree complex wavelet transform. IET Comput Vis. 2015 Apr 1;9(2):163\u0026ndash;73.\u003c/li\u003e\n \u003cli\u003eQazi AS, Farooq MS, Rustam F, Villar MG, Rodr\u0026iacute;guez CL, Ashraf I. Emotion Detection Using Facial Expression Involving Occlusions and Tilt. Appl Sci [Internet]. 2022 Nov 20 [cited 2024 May 31];12(22):11797. Available from: https://www.mdpi.com/2076-3417/12/22/11797/htm\u003c/li\u003e\n \u003cli\u003eValstar MF, Mehu M, Jiang B, Pantic M, Scherer K. Meta-analysis of the first facial expression recognition challenge. IEEE Trans Syst Man, Cybern Part B Cybern [Internet]. 2012 [cited 2024 Feb 29];42(4):966\u0026ndash;79. Available from: http://ieeexplore.ieee.org.\u003c/li\u003e\n \u003cli\u003eJia S, Wang S, Hu C, Webster PJ, Li X. Detection of Genuine and Posed Facial Expressions of Emotion: Databases and Methods [Internet]. Vol. 11, Frontiers in Psychology. Frontiers Media SA; 2021 [cited 2024 Feb 29]. p. 580287. Available from: www.frontiersin.org\u003c/li\u003e\n \u003cli\u003eXu X, Zheng W, Zong Y, Lu C, Jiang X. Sample Self-Revised Network for Cross-Dataset Facial Expression Recognition. In: Proceedings of the International Joint Conference on Neural Networks. Institute of Electrical and Electronics Engineers Inc.; 2022.\u003c/li\u003e\n \u003cli\u003eMeng H, Yuan F, Tian Y, Yan T. Cross-datasets facial expression recognition via distance metric learning and teacher-student model. Multimed Tools Appl [Internet]. 2022 [cited 2024 Feb 29];81(4):5621\u0026ndash;43. Available from: https://doi.org/10.1007/s11042-021-11765-4\u003c/li\u003e\n \u003cli\u003eLi S, Deng W. Deep Emotion Transfer Network for Cross-database Facial Expression Recognition. In: Proceedings - International Conference on Pattern Recognition. 2018. p. 3092\u0026ndash;9.\u003c/li\u003e\n \u003cli\u003eMayer C, Eggers M, Radig B. Cross-database evaluation for facial expression recognition. Pattern Recognit Image Anal [Internet]. 2014 [cited 2024 Mar 7];24(1):124\u0026ndash;32. Available from: https://sci-hub.se/10.1134/s1054661814010106\u003c/li\u003e\n \u003cli\u003eGu W, Xiang C, Venkatesh Y V., Huang D, Lin H. Facial expression recognition using radial encoding of local Gabor features and classifier synthesis. Pattern Recognit. 2012 Jan 1;45(1):80\u0026ndash;91.\u003c/li\u003e\n \u003cli\u003eAbd El Meguid MK, Levine MD. Fully automated recognition of spontaneous facial expressions in videos using random forest classifiers. IEEE Trans Affect Comput. 2014;5(2):141\u0026ndash;54.\u003c/li\u003e\n \u003cli\u003eShan C, Gong S, McOwan PW. Facial expression recognition based on Local Binary Patterns: A comprehensive study. Image Vis Comput [Internet]. 2009 May 4 [cited 2017 Aug 3];27(6):803\u0026ndash;16. Available from: http://www.sciencedirect.com/science/article/pii/S0262885608001844\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-5379043/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5379043/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eEmotion detection is a critical aspect of human-computer interaction and various other applications, including real-time video surveillance, psychological studies and social media emotion analysis. Understanding the complexity and nuance of human emotions requires an in-depth look at emotion intensity. Furthermore, occlusion is a common problem in facial emotion detection, leading to low detection accuracy. This paper presents a unique method for utilising Support Vector Machine (SVM) classification with wavelet coefficient analysis to analyse emotion intensities. To assess its effectiveness, we tested our approach on several benchmark datasets, such as the Japanese Female Facial Expression (JAFFE) dataset and the Facial Expression Recognition 2013 (FER2013) dataset. Our method demonstrates scalability and adaptability for application in large datasets, addressing the needs of big data environments. Additionally, we tested our method on occluded sections of the FER2013 dataset to simulate real-world conditions where facial emotions might be partially obscured, such as in video surveillance settings. Results indicate that wavelet coefficient analysis successfully captures nuanced differences in emotional intensity, while SVM performs robustly in classifying emotions, even in challenging, partially occluded conditions. Furthermore, the system\u0026rsquo;s structure allows for effective implementation in real-time applications and adaptability to the vast data streams typical of social media and large-scale psychological datasets. Our findings suggest that the proposed framework improves detection processes and offers a scalable, adaptable solution for large-scale emotion analysis across varied applications and significant data contexts.\u003c/p\u003e","manuscriptTitle":"Spontaneous vs Posed: Analysis of Emotion Intensities using Wavelet Coefficient and Support Vector Machine","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-11-19 10:55:56","doi":"10.21203/rs.3.rs-5379043/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"d2a77585-ada1-4772-8d23-5a1b6b1d0be6","owner":[],"postedDate":"November 19th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-09-13T06:38:19+00:00","versionOfRecord":[],"versionCreatedAt":"2024-11-19 10:55:56","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5379043","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5379043","identity":"rs-5379043","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}