Semantic-Aware Image Deduplication: Leveraging Object Recognition for Enhanced Accuracy

Semantic-Aware Image Deduplication: Leveraging Object Recognition for Enhanced Accuracy | 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 Semantic-Aware Image Deduplication: Leveraging Object Recognition for Enhanced Accuracy Rahul Shah, Ashok kumar Shrivastava This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6396148/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 01 Aug, 2025 Read the published version in International Journal of Intelligent Engineering and Systems → Version 1 posted You are reading this latest preprint version Abstract Image deduplication is a critical task in managing large-scale image databases, with applications ranging from content moderation to efficient storage solutions. Traditional methods often rely on pixel-level comparisons or perceptual hashing techniques, which may fail to capture semantic similarities between images. This paper proposes a novel approach to image deduplication that incorporates semantic information derived from object recognition algorithms. By leveraging state-of-the-art deep learning models for object detection and classification, our method achieves higher accuracy in identifying duplicate and near-duplicate images, even in cases where visual features may differ significantly. We present a comprehensive evaluation of our semantic-aware deduplication system on diverse datasets, demonstrating its superior performance compared to existing methods. Our findings suggest that integrating semantic understanding into image deduplication processes can significantly enhance accuracy and robustness, paving the way for more efficient and intelligent image management systems. Image Deduplication Object Recognition Semantic Similarity Deep Learnings Computer Vision Content-Based Image Retrieval Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction The exponential growth of digital images in recent years has presented significant challenges in data management, storage optimization, and content curation. Image deduplication, the process of identifying and managing duplicate or near-duplicate images within large datasets, has become increasingly important in various domains, including social media platforms, e-commerce websites, and digital asset management systems [1]. Traditional approaches to image deduplication have primarily focused on low-level visual features or perceptual hashing techniques [2,3]. While these methods have shown success in certain scenarios, they often struggle with semantically similar images that may differ in visual appearance due to factors such as lighting conditions, viewpoint changes, or minor edits [4]. This paper introduces a novel approach to image deduplication that incorporates semantic information derived from object recognition algorithms. By leveraging the power of deep learning models for object detection and classification, our method aims to capture higher-level semantic similarities between images, leading to more accurate and robust deduplication results. The primary contributions of this work are as follows: 1. A semantic-aware image deduplication framework that integrates object recognition techniques with traditional visual feature analysis. 2. A novel similarity metric that combines visual and semantic information to identify duplicate and near-duplicate images. 3. An extensive evaluation of the proposed method on diverse datasets, demonstrating its superior performance compared to existing deduplication techniques. 4. An analysis of the computational requirements and scalability of the semantic-aware approach in real-world applications. The remainder of this paper is organized as follows: Section 2 provides a comprehensive review of related work in image deduplication and object recognition. Section 3 describes the proposed semantic-aware deduplication framework in detail. Section 4 presents the experimental setup and evaluation methodology. Section 5 discusses the results and provides a comparative analysis with existing methods. Finally, Section 6 concludes the paper and outlines future research directions. 2. Related Work 2.1 Traditional Image Deduplication Techniques Several Image deduplication has been an active area of research for several decades, with various approaches proposed to address the challenge of identifying duplicate and near-duplicate images efficiently and accurately. Early methods focused on pixel-level comparisons, which were computationally expensive and sensitive to minor variations in image content [5]. Subsequent research introduced more sophisticated techniques based on image hashing and feature extraction. 2.1.1 Perceptual Hashing Perceptual hashing algorithms, such as pHash [6] and aHash [7], aim to generate compact representations of images that are resistant to minor modifications while preserving the overall visual structure. These methods typically involve reducing the image to a fixed size, converting it to grayscale, and computing a hash based on the discrete cosine transform (DCT) or average pixel values. While perceptual hashing techniques are computationally efficient and effective for detecting exact or near-exact duplicates, they may struggle with semantically similar images that differ significantly in visual appearance [8]. 2.1.2 Content Based Image Retrieval (CBIR) Perceptual Content-based image retrieval techniques extract low-level visual features such as color histograms, texture descriptors, and edge information to represent images [9]. These features are then used to compute similarity scores between images, allowing for the identification of duplicates and near-duplicates. Popular CBIR-based approaches include the use of local feature descriptors like SIFT (Scale-Invariant Feature Transform) [10] and SURF (Speeded up Robust Features) [11]. While CBIR methods offer improved robustness compared to simple hashing techniques, they may still struggle with capturing high-level semantic similarities between images [12]. 2.2 Deep Learning in Image Analysis The advent of deep learning has revolutionized the field of computer vision, leading to significant advancements in various image analysis tasks, including object detection, image classification, and semantic segmentation [13]. Convolutional Neural Networks (CNNs) have emerged as the dominant architecture for image-related tasks, demonstrating remarkable performance on benchmark datasets [14]. 2.2.1 Object Detection and Retention Object detection and recognition have seen substantial improvements with the introduction of deep learning models such as R-CNN [15], Fast R-CNN [16], and YOLO (You Only Look Once) [17]. These models can accurately identify and localize multiple objects within an image, providing rich semantic information about the image content. Recent advancements, such as Mask R-CNN [18] and EfficientDet [19], have further improved the accuracy and efficiency of object detection systems. 2.2.2 Image Similarity Using Deep Features Researchers have explored the use of deep learning features for image similarity tasks, including image retrieval and deduplication [20]. By extracting features from pre-trained CNN models, such as VGGNet [21] or ResNet [22], it is possible to create compact and semantically rich representations of images. These deep features have shown superior performance in capturing both low-level and high-level image characteristics compared to traditional handcrafted features [23]. 2.3 Semantic-Based Approaches in Image Analysis While deep learning has significantly improved the accuracy of image analysis tasks, there is growing interest in incorporating explicit semantic information to enhance performance further. Semantic-based approaches aim to bridge the gap between low-level visual features and high-level semantic concepts, leading to more robust and interpretable image analysis systems [24]. 2.3.1 Semantic Segmentation Semantic segmentation techniques, which assign class labels to each pixel in an image, have demonstrated the potential of combining visual and semantic information [25]. Models such as DeepLab [26] and PSPNet [27] have achieved state-of-the-art results on various semantic segmentation benchmarks, providing fine-grained semantic information about image content. 2.3.2 Visual Relationship Detection Recent work in visual relationship detection aims to identify not only objects within an image but also the relationships between them [28]. This higher-level semantic understanding can provide valuable context for image analysis tasks, including deduplication. Models such as VTransE [29] and Neural Motifs [30] have shown promising results in capturing complex semantic relationships within images. 2.4 Gap in Current Research Despite the advancements in both image deduplication and semantic image analysis, there remains a significant gap in leveraging semantic information for enhancing deduplication accuracy. Most existing deduplication methods still rely primarily on visual features or perceptual hashing, which may fail to capture important semantic similarities between images. While some recent works have explored the use of deep features for image similarity tasks [31,32], there is limited research on explicitly incorporating object recognition and semantic relationships into the deduplication process. This paper aims to address this gap by proposing a semantic-aware image deduplication framework that leverages state-of-the-art object recognition techniques to enhance the accuracy and robustness of duplicate detection. By combining visual and semantic information, our approach seeks to improve upon existing methods, particularly in cases where images are semantically similar but visually distinct. 3. Proposed Semantic Aware Deduplication Framework This section presents our novel semantic-aware image deduplication framework, which integrates object recognition techniques with traditional visual feature analysis to achieve higher accuracy in identifying duplicate and near-duplicate images. The proposed framework consists of several key components, as illustrated in Fig. 1 . 3.1 System Architecture The proposed framework comprises the following main components: 1. Image Pre-processing Module 2. Visual Feature Extraction Module 3. Object Recognition Module 4. Semantic Feature Extraction Module 5. Similarity Computation Module 6. Decision Module Each of these components plays a crucial role in the overall deduplication process, combining visual and semantic information to make accurate similarity assessments. 3.2 Image Pre-processing The image pre-processing module is responsible for preparing input images for subsequent analysis. This stage involves several steps: 1. Resizing: All input images are resized to a standard resolution (e.g., 224x224 pixels) to ensure consistent processing across different image sizes. 2. Color Space Conversion: Images are converted to a consistent color space (e.g., RGB) if necessary. 3. Normalization: Pixel values are normalized to a common range (e.g., [0, 1]) to improve the stability of feature extraction and object recognition processes. 4. Data Augmentation: For training purposes, data augmentation techniques such as random cropping, flipping, and color jittering may be applied to increase the robustness of the model. 3.3 Visual Feature Extraction To The visual feature extraction module employs a deep convolutional neural network (CNN) to extract low-level and mid-level visual features from the pre-processed images. We utilize a pre-trained CNN architecture, such as ResNet-50 [22] or Efficient Net [33], as the backbone for feature extraction. The output of this module is a high-dimensional feature vector that captures various visual characteristics of the image. 3.3.1 Transfer Learning and Fine-Tuning To adapt the pre-trained CNN to our specific deduplication task, we employ transfer learning techniques. The network is fine-tuned on a dataset of known duplicate and non-duplicate image pairs, allowing it to learn relevant features for deduplication. We use a Siamese network architecture [34] during fine-tuning to learn a similarity metric directly from image pairs. 3.4 Object Recognition The object recognition module is a critical component of our semantic-aware framework. It employs a state-of-the-art object detection and recognition model to identify and localize objects within the input images. We use a two-stage object detection approach, such as Faster R-CNN [35] or Mask R-CNN [18], pre-trained on a large-scale object detection dataset like COCO [36]. The object recognition module performs the following tasks: 1. Object Detection: Identifies bounding boxes for potential objects in the image. 2. Object Classification: Assigns class labels to detected objects with associated confidence scores. 3. Object Localization: Provides precise spatial information for each detected object The output of this module is a list of detected objects, their class labels, confidence scores, and bounding box coordinates. 3.5 Object Recognition The semantic feature extraction module processes the output of the object recognition module to generate a semantic representation of the image. This representation captures high-level information about the objects present in the image, their relationships, and their spatial arrangement. 3.5.1 Object-based Semantic Embedding We propose an object-based semantic embedding that encodes the detected objects and their attributes into a fixed-length vector. This embedding includes: 1. Object Class Distribution: A histogram of object classes present in the image, weighted by their confidence scores and relative sizes. 2. Spatial Relationships: Encoding of the relative positions and sizes of objects using a spatial pyramid representation [37]. 3. Attribute Information: Incorporation of relevant object attributes (e.g., color, texture) when available from the object recognition model. 3.5.2 Graph-based Semantic Representation To capture more complex relationships between objects, we also propose a graph-based semantic representation. In this approach, detected objects are represented as nodes in a graph, with edges representing spatial and semantic relationships between objects. We use a Graph Convolutional Network (GCN) [38] to learn a compact representation of this object relationship graph. 3.6 Similarity Computation The similarity computation module combines the visual and semantic features to calculate a comprehensive similarity score between image pairs. We propose a multi-modal similarity metric that takes into account both visual and semantic similarities: S(I1, I2) = α * Sv(V1, V2) + (1 - α) * Ss(S1, S2) Where: ● S(I1, I2) is the overall similarity score between images I1 and I2 ● Sv(V1, V2) is the visual similarity based on the extracted visual features ● Ss(S1, S2) is the semantic similarity based on the semantic representations ● α is a weighting parameter that balances the contribution of visual and semantic similarities. 3.6.1 Visual Similarity The visual similarity Sv is computed using a distance metric (e.g., cosine similarity or Euclidean distance) between the visual feature vectors extracted by the CNN. 3.6.2 Semantic Similarity The semantic similarity Ss is calculated using a combination of techniques: 1. Object Set Similarity: Jaccard similarity between the sets of detected objects in the two images. 2. Semantic Embedding Similarity: Cosine similarity between the object-based semantic embeddings. 3. Graph Similarity: If using the graph-based representation, we employ graph matching techniques or graph kernels to compute similarity between the object relationship graphs. 3.7 Decision Module The decision module determines whether two images are duplicates or near-duplicates based on the computed similarity score. We employ a threshold-based approach, where image pairs with a similarity score above a certain threshold are classified as duplicates. The threshold is determined through empirical analysis on a validation dataset, balancing precision and recall. 3.7.1 Adaptive Thresholding To account for variations in different image domains and use cases, we implement an adaptive thresholding mechanism. This approach adjusts the decision threshold based on the characteristics of the input images and the specific requirements of the application (e.g., prioritizing precision over recall or vice versa). 3.7.2 Confidence Estimation In addition to the binary duplicate/non-duplicate decision, our framework provides a confidence estimate for each classification. This confidence score is derived from the distance between the computed similarity score and the decision threshold, allowing for more nuanced decision-making in borderline cases. 3.8 Training and Optimization The entire framework is trained end-to-end using a large dataset of labeled image pairs (duplicate and non-duplicate). We employ a multi-task learning approach, optimizing for both visual and semantic similarity simultaneously. The loss function combines: 1. Contrastive loss for visual similarity learning 2. Cross-entropy loss for object recognition 3. Graph-based loss for semantic relationship learning Optimization is performed using stochastic gradient descent with momentum or adaptive learning rate methods such as Adam [39]. 4. Experimental Setup and Evaluation Methodology To evaluate the effectiveness of our proposed semantic-aware image deduplication framework, we conducted a comprehensive set of experiments on multiple datasets. This section details our experimental setup, including datasets, evaluation metrics, and baseline methods for comparison. 4.1 Datasets We utilized the following datasets in our experiments: 1. MNIST-Duplicate: A synthetic dataset created by applying random transformations (rotation, scaling, translation) to MNIST digits, resulting in 100,000 image pairs (50,000 duplicate pairs and 50,000 non-duplicate pairs). 2. Oxford Buildings Dataset: A real-world dataset containing 5,062 images of Oxford landmarks, with ground truth information on duplicate and near-duplicate images [40]. 3. Web Image Dataset (WID): A large-scale dataset of 1 million web images, with known duplicate and near-duplicate pairs, collected from various online sources [41]. 4. COCO-Duplicate: A custom dataset created from the COCO dataset [36] by applying various transformations and modifications to generate duplicate and near-duplicate pairs, focusing on images with multiple objects. Table 1 Overview of datasets used in the experiments Dataset Images Duplicate Pairs Non-Duplicate Pairs Object Types Transformation Types MNIST-Duplicate 200,000 50,000 50,000 Digits Rotation, Scaling, Translation Oxford Buildings 5,062 2,531 2,531 Buildings, Landmarks Viewpoint changes, Lighting variations Web Image Dataset (WID) 1,000,000 100,000 900,000 Various Multiple (natural variations) COCO-Duplicate 200,000 50,000 50,000 Multiple object categories Cropping, Color adjustments, Object addition/removal 4.2 Evaluation Metrics To assess the performance of our semantic-aware image deduplication framework, we employed the following evaluation metrics: 1. Precision: The ratio of correctly identified duplicate pairs to the total number of pairs identified as duplicates. 2. Recall: The ratio of correctly identified duplicate pairs to the total number of actual duplicate pairs in the dataset. 3. F1-score: The harmonic mean of precision and recall, providing a balanced measure of the model's performance. 4. Mean Average Precision (mAP): A metric that considers the ranking of duplicate images, particularly useful for evaluating near-duplicate detection. 5. Receiver Operating Characteristic (ROC) curve: A plot of the true positive rate against the false positive rate at various threshold settings. 6. Area Under the ROC Curve (AUC): A single scalar value representing the overall performance across all possible thresholds. 4.3 Baseline Methods We compared our semantic-aware framework against the following baseline methods: 1. Perceptual Hashing (pHash) [6]: A classic image hashing technique based on the discrete cosine transform. 2. SIFT + BoVW: Scale-Invariant Feature Transform descriptors with Bag of Visual Words representation [10]. 3. Deep Siamese Network: A convolutional neural network trained with a contrastive loss function [34]. 4. DeepRank: A deep learning-based approach for image similarity ranking [42]. 5. DupNet: A recent deep learning method specifically designed for image deduplication [43]. 4.4 Implementation Details Our semantic-aware image deduplication framework was implemented using PyTorch 1.8.0. The visual feature extraction module utilized a ResNet-50 architecture pre-trained on ImageNet, while the object recognition module employed a Faster R-CNN model pre-trained on COCO. The semantic feature extraction module used a Graph Convolutional Network with 3 layers. Training was performed on 4 NVIDIA Tesla V100 GPUs with 32GB memory each. We used the Adam optimizer with an initial learning rate of 1e-4 and a batch size of 64 image pairs. The model was trained for 50 epochs, with learning rate decay by a factor of 0.1 every 20 epochs. 4.5 Experimental Protocol We conducted our experiments following this protocol: 1. Data Preparation: Each dataset was split into training (60%), validation (20%), and test (20%) sets. Care was taken to ensure that duplicate pairs were not split across sets. 2. Model Training: The semantic-aware framework and baseline models were trained on the training set, with hyperparameters tuned using the validation set. 3. Threshold Selection: For each method, we determined the optimal decision threshold using the validation set, maximizing the F1-score. 4. Performance Evaluation: We evaluated all methods on the test set using the selected thresholds and computed the metrics described in Section 4.2. 5. Ablation Studies: We conducted ablation studies to analyze the impact of different components of our framework, including: ○ Visual features only vs. Semantic features only vs. Combined features ○ Impact of different object recognition models ○ Contribution of graph-based semantic representation 6. Scalability Analysis: We assessed the computational requirements and scalability of our approach by measuring processing time and memory usage on datasets of increasing size. 4.6 Statistical Analysis To ensure the statistical significance of our results, we performed the following analyses: 1. Confidence Intervals: We computed 95% confidence intervals for all reported metrics using bootstrap resampling with 1000 iterations. 2. Paired t-tests: We conducted paired t-tests to compare the performance of our method against each baseline, with a significance level of α = 0.05. 3. McNemar's Test: For binary classification decisions (duplicate vs. non-duplicate), we used McNemar's test to assess the statistical significance of improvements over baseline methods. 5. Results and Discussion In this section, we present the results of our experimental evaluation and provide a detailed analysis of the performance of our semantic-aware image deduplication framework compared to baseline methods. 5.1 Overall Performance Comparison Figure 2 shows the performance of our proposed method and baseline approaches across all datasets, reporting precision, recall, F1-score, and MAP. Our semantic-aware framework consistently outperforms all baseline methods across all datasets, achieving significant improvements in both F1-score and mAP. The performance gain is particularly notable on the COCO-Duplicate dataset, which contains images with multiple objects and complex semantic relationships. 5.2 ROC Curve Analysis Figure 3 presents the ROC curves for all methods on the Web Image Dataset (WID), illustrating the trade-off between true positive rate and false positive rate across different decision thresholds. The ROC curve analysis reveals that our semantic-aware approach maintains a higher true positive rate across all false positive rate levels, indicating superior performance in distinguishing between duplicate and non-duplicate images. The Area under the Curve (AUC) values are reported in Table 2 . Table 2 AUC values for different methods on the Web Image Dataset Method AUC pHash 0.867 SIFT + BoVW 0.901 Deep Siamese 0.953 DeepRank 0.961 DupNet 0.968 Ours (Semantic-Aware) 0.984 5.3 Ablation Studies To understand the contribution of different components in our framework, we conducted ablation studies. Table 4 presents the results of these studies on the COCO-Duplicate dataset. Table 3 Ablation study results on COCO-Duplicate dataset Method AUC mAP Visual features only 0.891 ± 0.005 0.878 Semantic features only 0.903 ± 0.004 0.891 Combined (no graph) 0.918 ± 0.003 0.909 Full model (with graph) 0.929 ± 0.003 0.921 The ablation studies demonstrate that both visual and semantic features contribute significantly to the overall performance. The integration of graph-based semantic representation further enhances the accuracy of our framework. 5.4 Impact of Object Recognition Models We experimented with different object recognition models to assess their impact on the framework's performance. Table 5 shows the results using various object detection architectures on the COCO-Duplicate dataset. Table 4 Performance with different object detection models on COCO-Duplicate dataset. Object Detection Model F1-score mAP Faster R-CNN 0.929 ± 0.003 0.921 Mask R-CNN 0.932 ± 0.003 0.924 YOLO v4 0.925 ± 0.004 0.917 EfficientDet 0.930 ± 0.003 0.922 While all object detection models provide strong performance, Mask R-CNN shows a slight edge, likely due to its ability to provide more detailed object segmentation information. 5.5 Scalability Analysis To assess the scalability of our approach, we measured the processing time and memory usage for deduplication tasks of increasing dataset sizes. Figure 3 illustrates the relationship between dataset size and computational resources required. Our semantic-aware framework shows a linear increase in processing time and memory usage as the dataset size grows, indicating good scalability for large-scale deduplication tasks. However, it does require more computational resources compared to simpler methods like perceptual hashing. 5.6 Qualitative Analysis Our semantic-aware approach demonstrates superior performance in handling cases where images are semantically similar but visually distinct, such as different views of the same landmark or products with varying backgrounds. 5.7 Discussion The experimental results clearly demonstrate the effectiveness of our semantic-aware image deduplication framework. Key findings include: 1. Consistent Performance Improvement: Our approach consistently outperforms existing methods across all datasets, with statistically significant improvements in F1-score and mAP. 2. Robustness to Visual Variations: The semantic-aware framework shows particular strength in handling cases where duplicate images have significant visual differences but share semantic content. 3. Complementary Features: The ablation studies confirm that both visual and semantic features contribute to the overall performance, with the graph-based semantic representation providing additional gains. 4. Scalability Trade-offs: While our method requires more computational resources than simpler approaches, it scales linearly with dataset size, making it feasible for large-scale applications. 5. Flexibility: The framework's performance can be further optimized by choosing appropriate object detection models for specific use cases or domains. Despite these promising results, there are some limitations and areas for future work: 1. Computational Overhead: The inclusion of object recognition and graph-based processing increases computational requirements, which may be a concern for extremely large-scale or real-time applications. 2. Domain Specificity: While our approach generalizes well across the tested datasets, further investigation is needed to assess its performance on highly specialized image collections (e.g., medical imaging, satellite imagery). 3. Fine-grained Similarity: Future work could explore more nuanced similarity metrics that consider the relative importance of different objects or semantic relationships within images. In conclusion, our semantic-aware image deduplication framework represents a significant advancement in the field, effectively leveraging both visual and semantic information to achieve state-of-the-art performance in identifying duplicate and near-duplicate images. 6. Conclusion and future work This paper presented a novel semantic-aware image deduplication framework that leverages object recognition techniques to enhance the accuracy of duplicate and near-duplicate image detection. Our approach integrates visual feature analysis with semantic information derived from state-of-the-art object detection models, demonstrating significant improvements over existing methods across diverse datasets. Key contributions of this work include: 1. A comprehensive framework that combines visual and semantic features for image deduplication. 2. A novel graph-based semantic representation that captures complex relationships between objects in images. 3. Extensive experimental evaluation demonstrating the effectiveness and scalability of the proposed approach. The results of our experiments highlight the potential of semantic-aware techniques in addressing the limitations of traditional image deduplication methods, particularly in scenarios involving visually distinct but semantically similar images. Future research directions include: 1. Exploring more advanced semantic representations: Investigating the use of scene graphs or visual relationship detection to capture even more nuanced semantic information. 2. Incorporating temporal information: Extending the framework to handle video deduplication by considering temporal relationships between frames. 3. Adaptive feature weighting: Developing mechanisms to dynamically adjust the importance of visual and semantic features based on the characteristics of the input images. 4. Privacy-preserving deduplication: Investigating techniques to perform semantic-aware deduplication while preserving user privacy, especially for sensitive applications. 5. Cross-modal deduplication: Extending the framework to identify duplicates across different modalities, such as images and text descriptions. 6. Unsupervised and self-supervised learning: Exploring ways to reduce the reliance on large labeled datasets for training the deduplication model. 7. Edge deployment: Optimizing the framework for deployment on edge devices with limited computational resources. In conclusion, semantic-aware image deduplication represents a promising direction for improving the accuracy and robustness of duplicate detection systems. As the volume and complexity of digital images continue to grow, such advanced techniques will become increasingly important for efficient data management and content curation across various domains. Declarations Conflicts of interest The authors declare that they have no conflict of interests. Ethics, Consent to Participate, and Consent to Publish declarations Not applicable. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Author Contribution R.S. (Rahul Shah) conceptualized the study and designed the research framework. R.S. and A.K.S. (Dr. Ashok Kumar Shrivastava) developed the methodology and conducted the experiments. A.K.S. contributed to data collection and preprocessing. R.S. performed the analysis and interpretation of results. R.S. wrote the main manuscript text, and A.K.S. reviewed and provided critical revisions. Both authors approved the final version of the manuscript. Acknowledgment We would like to express our sincere gratitude to everyone who contributed to this research. Special thanks to our colleagues for their insightful feedback and support throughout the research. We are also grateful to the institutions and organizations that provided the resources and datasets necessary for this study. Finally, we thank our families and friends Data Availability The data supporting the findings of this study are available from the corresponding author upon reasonable request. Due to certain restrictions, if applicable, such as privacy, confidentiality, or proprietary concerns, the dataset is not publicly accessible. References Wang, X., et al. A survey on visual content-based video indexing and retrieval. *IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews)*. 2014; 41(6): 797–819. Chum, O., et al. (2008). Near duplicate image detection: min-hash and tf-idf weighting. In BMVC (Vol. 810, pp. 812–815). Zauner, C. (2010). Implementation and benchmarking of perceptual image hash functions. Master's thesis, University of Applied Sciences Hagenberg, Austria. Ke, Y., et al. (2004). Efficient near-duplicate detection and sub-image retrieval. In Proceedings of the 12th annual ACM international conference on Multimedia (pp. 869–876). Huang, J., et al. (2011). Image indexing using color correlograms. In 2011 IEEE Conference on Computer Vision and Pattern Recognition (pp. 762–768). Monga, V., & Evans, B. L. (2006). Perceptual image hashing via feature points: performance evaluation and tradeoffs. IEEE transactions on Image Processing, 15(11), 3452–3465. Yang, B., et al. (2006). A comparative study on robust hash functions for digital audio. In International Workshop on Digital Watermarking (pp. 354–365). Springer, Berlin, Heidelberg. Jégou, H., et al. (2008). Hamming embedding and weak geometric consistency for large scale image search. In European conference on computer vision (pp. 304–317). Springer, Berlin, Heidelberg. Smeulders, A. W., et al. (2000). Content-based image retrieval at the end of the early years. IEEE Transactions on pattern analysis and machine intelligence, 22(12), 1349–1380. Lowe, D. G. (2004). Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2), 91–110. Bay, H., et al. (2008). Speeded-up robust features (SURF). Computer vision and image understanding, 110(3), 346–359. Liu, Y., et al. (2007). A survey of content-based image retrieval with high-level semantics. Pattern recognition, 40(1), 262–282. LeCun, Y., et al. (2015). Deep learning. nature, 521(7553), 436–444. Krizhevsky, A., et al. (2012). Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 1097–1105. Girshick, R., et al. (2014). Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 580–587). Girshick, R. (2015). Fast r-cnn. In Proceedings of the IEEE international conference on computer vision (pp. 1440–1448). Redmon, J., et al. (2016). You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 779–788). He, K., et al. (2017). Mask r-cnn. In Proceedings of the IEEE international conference on computer vision (pp. 2961–2969). Tan, M., et al. (2020). Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 10781–10790). Gordo, A., et al. (2016). Deep image retrieval: Learning global representations for image search. In European conference on computer vision (pp. 241–257). Springer, Cham. Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556. He, K., et al. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770–778). Babenko, A., & Lempitsky, V. (2015). Aggregating local deep features for image retrieval. In Proceedings of the IEEE international conference on computer vision (pp. 1269–1277). Zhang, H., et al. (2018). Visual semantic reasoning for image-text matching. In Proceedings of the IEEE International Conference on Computer Vision (pp. 4654–4662). Long, J., et al. (2015). Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 3431–3440). Chen, L. C., et al. (2017). Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4), 834–848. Zhao, H., et al. (2017). Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2881–2890). Lu, C., et al. (2016). Visual relationship detection with language priors. In European conference on computer vision (pp. 852–869). Springer, Cham. Zhang, H., et al. (2017). Visual translation embedding network for visual relation detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 5532–5540). Zellers, R., et al. (2018). Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 5831–5840). Zheng, L., et al. (2017). SIFT meets CNN: A decade survey of instance retrieval. IEEE transactions on pattern analysis and machine intelligence, 40(5), 1224–1244. Yang, J., et al. (2017). Stacked attention networks for image question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 21–29). Tan, M., & Le, Q. (2019). Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning (pp. 6105–6114). PMLR. Koch, G., et al. (2015). Siamese neural networks for one-shot image recognition. In ICML deep learning workshop (Vol. 2). Ren, S., et al. (2015). Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 91–99. Lin, T. Y., et al. (2014). Microsoft coco: Common objects in context. In European conference on computer vision (pp. 740–755). Springer, Cham. Lazebnik, S., et al. (2006). Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06) (Vol. 2, pp. 2169–2178). IEEE. Kipf, T. N., & Welling, M. (2016). Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907. Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. Philbin, J., et al. (2007). Object retrieval with large vocabularies and fast spatial matching. In 2007 IEEE conference on computer vision and pattern recognition (pp. 1–8). IEEE. Wang, X. J., et al. (2013). Duplicate discovery on 2 billion internet images. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 429–436). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 01 Aug, 2025 Read the published version in International Journal of Intelligent Engineering and Systems → 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-6396148","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":462508614,"identity":"bfeb504e-0407-4e66-b075-d3eef534515b","order_by":0,"name":"Rahul Shah","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+ElEQVRIiWNgGAWjYFACHgaGBwwMCQwSjI0PQHw2BgSJW0sCWAtzswFYMRvxWtjbJBAW4NFi3n72mEQCQ10e/+zGtsofDHbRfPI9Bgwfyg7j1CJzJi8NqOVwscSdg223eRiSc9vYeAwYZ5zDrUWCIccMqOVAYsONxLbbDAzMYC3MvG14tPC/AWmpS5wP1FL4g6EeouUvPi0SYFuYEzcAtQCD4jBECyNeLW+MLRIMDiduvJHYLM1jcByoJa3gYM+5dDwOyzG88aGiLnHejfSHH39UVOfObz688cGPMmucWiDAAIXBYXCAgHoMwP6AVB2jYBSMglEwvAEAe2RRE6YDoeMAAAAASUVORK5CYII=","orcid":"","institution":"Amity University Madhya Pradesh","correspondingAuthor":true,"prefix":"","firstName":"Rahul","middleName":"","lastName":"Shah","suffix":""},{"id":462508615,"identity":"bbf7aeff-d1ce-4d30-94f5-7915541a5666","order_by":1,"name":"Ashok kumar Shrivastava","email":"","orcid":"","institution":"Amity University Madhya Pradesh","correspondingAuthor":false,"prefix":"","firstName":"Ashok","middleName":"kumar","lastName":"Shrivastava","suffix":""}],"badges":[],"createdAt":"2025-04-07 16:23:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6396148/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6396148/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.22266/ijies2025.0831.09","type":"published","date":"2025-08-02T00:00:00+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":83675570,"identity":"2655e662-d267-44fe-a9cf-6a0681809230","added_by":"auto","created_at":"2025-05-30 14:48:48","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":28602,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of the Semantic-Aware Image Deduplication Framework\u003c/p\u003e","description":"","filename":"Picture1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6396148/v1/c1e89fa5656a28f80a67ae57.jpg"},{"id":83676495,"identity":"809e25a4-4265-4ab8-9608-e4f7d0e9b38f","added_by":"auto","created_at":"2025-05-30 14:56:48","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":35446,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance comparison of different methods across datasets (mean ± 95% CI)\u003c/p\u003e","description":"","filename":"Picture2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6396148/v1/0e50147db93e27de86f1ac71.jpg"},{"id":83675576,"identity":"df095198-90c4-481e-9322-fb891f3e291e","added_by":"auto","created_at":"2025-05-30 14:48:48","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":11089,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves for different methods on the Web Image Dataset\u003c/p\u003e","description":"","filename":"Picture3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6396148/v1/b649ed05ee3b4fe1d45df377.jpg"},{"id":83675573,"identity":"03ff7f41-3243-4749-94a8-08eaf2bd345e","added_by":"auto","created_at":"2025-05-30 14:48:48","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":7319,"visible":true,"origin":"","legend":"\u003cp\u003eScalability analysis - Processing time and memory usage vs. dataset size\u003c/p\u003e","description":"","filename":"Picture4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6396148/v1/637df4996d57c5ee1f02efe3.jpg"},{"id":89599867,"identity":"e2d39eca-8074-4fc4-9840-96e4c1c6a42c","added_by":"auto","created_at":"2025-08-21 17:50:59","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1186906,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6396148/v1/24bc60c2-5df9-44eb-9c88-f34a6e350e83.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Semantic-Aware Image Deduplication: Leveraging Object Recognition for Enhanced Accuracy","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThe exponential growth of digital images in recent years has presented significant challenges in data management, storage optimization, and content curation. Image deduplication, the process of identifying and managing duplicate or near-duplicate images within large datasets, has become increasingly important in various domains, including social media platforms, e-commerce websites, and digital asset management systems [1]. Traditional approaches to image deduplication have primarily focused on low-level visual features or perceptual hashing techniques [2,3]. While these methods have shown success in certain scenarios, they often struggle with semantically similar images that may differ in visual appearance due to factors such as lighting conditions, viewpoint changes, or minor edits [4].\u003c/p\u003e\n\u003cp\u003eThis paper introduces a novel approach to image deduplication that incorporates semantic information derived from object recognition algorithms. By leveraging the power of deep learning models for object detection and classification, our method aims to capture higher-level semantic similarities between images, leading to more accurate and robust deduplication results. The primary contributions of this work are as follows:\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e1. A semantic-aware image deduplication framework that integrates object recognition techniques with traditional visual feature analysis.\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e2. A novel similarity metric that combines visual and semantic information to identify duplicate and near-duplicate images.\u003c/p\u003e\n\u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e3. An extensive evaluation of the proposed method on diverse datasets, demonstrating its superior performance compared to existing deduplication techniques.\u003c/p\u003e\n\u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e4. An analysis of the computational requirements and scalability of the semantic-aware approach in real-world applications.\u003c/p\u003e\n\u003c/span\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eThe remainder of this paper is organized as follows: Section 2 provides a comprehensive review of related work in image deduplication and object recognition. Section 3 describes the proposed semantic-aware deduplication framework in detail. Section 4 presents the experimental setup and evaluation methodology. Section 5 discusses the results and provides a comparative analysis with existing methods. Finally, Section 6 concludes the paper and outlines future research directions.\u003c/p\u003e"},{"header":"2. Related Work","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Traditional Image Deduplication Techniques\u003c/h2\u003e \u003cp\u003eSeveral Image deduplication has been an active area of research for several decades, with various approaches proposed to address the challenge of identifying duplicate and near-duplicate images efficiently and accurately. Early methods focused on pixel-level comparisons, which were computationally expensive and sensitive to minor variations in image content [5]. Subsequent research introduced more sophisticated techniques based on image hashing and feature extraction.\u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section3\"\u003e \u003ch2\u003e2.1.1 Perceptual Hashing\u003c/h2\u003e \u003cp\u003ePerceptual hashing algorithms, such as pHash [6] and aHash [7], aim to generate compact representations of images that are resistant to minor modifications while preserving the overall visual structure. These methods typically involve reducing the image to a fixed size, converting it to grayscale, and computing a hash based on the discrete cosine transform (DCT) or average pixel values. While perceptual hashing techniques are computationally efficient and effective for detecting exact or near-exact duplicates, they may struggle with semantically similar images that differ significantly in visual appearance [8].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e2.1.2 Content Based Image Retrieval (CBIR)\u003c/h2\u003e \u003cp\u003ePerceptual Content-based image retrieval techniques extract low-level visual features such as color histograms, texture descriptors, and edge information to represent images [9]. These features are then used to compute similarity scores between images, allowing for the identification of duplicates and near-duplicates. Popular CBIR-based approaches include the use of local feature descriptors like SIFT (Scale-Invariant Feature Transform) [10] and SURF (Speeded up Robust Features) [11]. While CBIR methods offer improved robustness compared to simple hashing techniques, they may still struggle with capturing high-level semantic similarities between images [12].\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Deep Learning in Image Analysis\u003c/h2\u003e \u003cp\u003eThe advent of deep learning has revolutionized the field of computer vision, leading to significant advancements in various image analysis tasks, including object detection, image classification, and semantic segmentation [13]. Convolutional Neural Networks (CNNs) have emerged as the dominant architecture for image-related tasks, demonstrating remarkable performance on benchmark datasets [14].\u003c/p\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.2.1 Object Detection and Retention\u003c/h2\u003e \u003cp\u003eObject detection and recognition have seen substantial improvements with the introduction of deep learning models such as R-CNN [15], Fast R-CNN [16], and YOLO (You Only Look Once) [17]. These models can accurately identify and localize multiple objects within an image, providing rich semantic information about the image content. Recent advancements, such as Mask R-CNN [18] and EfficientDet [19], have further improved the accuracy and efficiency of object detection systems.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.2.2 Image Similarity Using Deep Features\u003c/h2\u003e \u003cp\u003eResearchers have explored the use of deep learning features for image similarity tasks, including image retrieval and deduplication [20]. By extracting features from pre-trained CNN models, such as VGGNet [21] or ResNet [22], it is possible to create compact and semantically rich representations of images. These deep features have shown superior performance in capturing both low-level and high-level image characteristics compared to traditional handcrafted features [23].\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Semantic-Based Approaches in Image Analysis\u003c/h2\u003e \u003cp\u003eWhile deep learning has significantly improved the accuracy of image analysis tasks, there is growing interest in incorporating explicit semantic information to enhance performance further. Semantic-based approaches aim to bridge the gap between low-level visual features and high-level semantic concepts, leading to more robust and interpretable image analysis systems [24].\u003c/p\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e2.3.1 Semantic Segmentation\u003c/h2\u003e \u003cp\u003eSemantic segmentation techniques, which assign class labels to each pixel in an image, have demonstrated the potential of combining visual and semantic information [25]. Models such as DeepLab [26] and PSPNet [27] have achieved state-of-the-art results on various semantic segmentation benchmarks, providing fine-grained semantic information about image content.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003e2.3.2 Visual Relationship Detection\u003c/h2\u003e \u003cp\u003eRecent work in visual relationship detection aims to identify not only objects within an image but also the relationships between them [28]. This higher-level semantic understanding can provide valuable context for image analysis tasks, including deduplication. Models such as VTransE [29] and Neural Motifs [30] have shown promising results in capturing complex semantic relationships within images.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Gap in Current Research\u003c/h2\u003e \u003cp\u003eDespite the advancements in both image deduplication and semantic image analysis, there remains a significant gap in leveraging semantic information for enhancing deduplication accuracy. Most existing deduplication methods still rely primarily on visual features or perceptual hashing, which may fail to capture important semantic similarities between images. While some recent works have explored the use of deep features for image similarity tasks [31,32], there is limited research on explicitly incorporating object recognition and semantic relationships into the deduplication process.\u003c/p\u003e \u003cp\u003eThis paper aims to address this gap by proposing a semantic-aware image deduplication framework that leverages state-of-the-art object recognition techniques to enhance the accuracy and robustness of duplicate detection. By combining visual and semantic information, our approach seeks to improve upon existing methods, particularly in cases where images are semantically similar but visually distinct.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Proposed Semantic Aware Deduplication Framework","content":"\u003cp\u003eThis section presents our novel semantic-aware image deduplication framework, which integrates object recognition techniques with traditional visual feature analysis to achieve higher accuracy in identifying duplicate and near-duplicate images. The proposed framework consists of several key components, as illustrated in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\n\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1 System Architecture\u003c/h2\u003e\n \u003cp\u003eThe proposed framework comprises the following main components:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Image Pre-processing Module\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Visual Feature Extraction Module\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Object Recognition Module\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e4. Semantic Feature Extraction Module\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e5. Similarity Computation Module\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e6. Decision Module\u003c/p\u003e\n \u003c/span\u003e\n \u003cp\u003eEach of these components plays a crucial role in the overall deduplication process, combining visual and semantic information to make accurate similarity assessments.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2 Image Pre-processing\u003c/h2\u003e\n \u003cp\u003eThe image pre-processing module is responsible for preparing input images for subsequent analysis. This stage involves several steps:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Resizing: All input images are resized to a standard resolution (e.g., 224x224 pixels) to ensure consistent processing across different image sizes.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Color Space Conversion: Images are converted to a consistent color space (e.g., RGB) if necessary.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Normalization: Pixel values are normalized to a common range (e.g., [0, 1]) to improve the stability of feature extraction and object recognition processes.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e4. Data Augmentation: For training purposes, data augmentation techniques such as random cropping, flipping, and color jittering may be applied to increase the robustness of the model.\u003c/p\u003e\n \u003c/span\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\n \u003ch2\u003e3.3 Visual Feature Extraction\u003c/h2\u003e\n \u003cp\u003eTo The visual feature extraction module employs a deep convolutional neural network (CNN) to extract low-level and mid-level visual features from the pre-processed images. We utilize a pre-trained CNN architecture, such as ResNet-50 [22] or Efficient Net [33], as the backbone for feature extraction. The output of this module is a high-dimensional feature vector that captures various visual characteristics of the image.\u003c/p\u003e\n \u003cdiv id=\"Sec17\" class=\"Section3\"\u003e\n \u003ch2\u003e3.3.1 Transfer Learning and Fine-Tuning\u003c/h2\u003e\n \u003cp\u003eTo adapt the pre-trained CNN to our specific deduplication task, we employ transfer learning techniques. The network is fine-tuned on a dataset of known duplicate and non-duplicate image pairs, allowing it to learn relevant features for deduplication. We use a Siamese network architecture [34] during fine-tuning to learn a similarity metric directly from image pairs.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\n \u003ch2\u003e3.4 Object Recognition\u003c/h2\u003e\n \u003cp\u003eThe object recognition module is a critical component of our semantic-aware framework. It employs a state-of-the-art object detection and recognition model to identify and localize objects within the input images. We use a two-stage object detection approach, such as Faster R-CNN [35] or Mask R-CNN [18], pre-trained on a large-scale object detection dataset like COCO [36].\u003c/p\u003e\n \u003cp\u003eThe object recognition module performs the following tasks:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Object Detection: Identifies bounding boxes for potential objects in the image.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Object Classification: Assigns class labels to detected objects with associated confidence scores.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Object Localization: Provides precise spatial information for each detected object\u003c/p\u003e\n \u003c/span\u003eThe output of this module is a list of detected objects, their class labels, confidence scores, and bounding box coordinates.\n\u003c/div\u003e\n\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e\n \u003ch2\u003e3.5 Object Recognition\u003c/h2\u003e\n \u003cp\u003eThe semantic feature extraction module processes the output of the object recognition module to generate a semantic representation of the image. This representation captures high-level information about the objects present in the image, their relationships, and their spatial arrangement.\u003c/p\u003e\n \u003cdiv id=\"Sec20\" class=\"Section3\"\u003e\n \u003ch2\u003e3.5.1 Object-based Semantic Embedding\u003c/h2\u003e\n \u003cp\u003eWe propose an object-based semantic embedding that encodes the detected objects and their attributes into a fixed-length vector. This embedding includes:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Object Class Distribution: A histogram of object classes present in the image, weighted by their confidence scores and relative sizes.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Spatial Relationships: Encoding of the relative positions and sizes of objects using a spatial pyramid representation [37].\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Attribute Information: Incorporation of relevant object attributes (e.g., color, texture) when available from the object recognition model.\u003c/p\u003e\n \u003c/span\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec21\" class=\"Section3\"\u003e\n \u003ch2\u003e3.5.2 Graph-based Semantic Representation\u003c/h2\u003e\n \u003cp\u003eTo capture more complex relationships between objects, we also propose a graph-based semantic representation. In this approach, detected objects are represented as nodes in a graph, with edges representing spatial and semantic relationships between objects. We use a Graph Convolutional Network (GCN) [38] to learn a compact representation of this object relationship graph.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e\n \u003ch2\u003e3.6 Similarity Computation\u003c/h2\u003e\n \u003cp\u003eThe similarity computation module combines the visual and semantic features to calculate a comprehensive similarity score between image pairs. We propose a multi-modal similarity metric that takes into account both visual and semantic similarities:\u003c/p\u003e\n \u003cp\u003eS(I1, I2) = \u0026alpha; * Sv(V1, V2) + (1 - \u0026alpha;) * Ss(S1, S2)\u003c/p\u003e\n \u003cp\u003eWhere:\u003c/p\u003e\n \u003cp\u003e● S(I1, I2) is the overall similarity score between images I1 and I2\u003c/p\u003e\n \u003cp\u003e● Sv(V1, V2) is the visual similarity based on the extracted visual features\u003c/p\u003e\n \u003cp\u003e● Ss(S1, S2) is the semantic similarity based on the semantic representations\u003c/p\u003e\n \u003cp\u003e● \u0026alpha; is a weighting parameter that balances the contribution of visual and semantic similarities.\u003c/p\u003e\n \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e\n \u003ch2\u003e3.6.1 Visual Similarity\u003c/h2\u003e\n \u003cp\u003eThe visual similarity Sv is computed using a distance metric (e.g., cosine similarity or Euclidean distance) between the visual feature vectors extracted by the CNN.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec24\" class=\"Section3\"\u003e\n \u003ch2\u003e3.6.2 Semantic Similarity\u003c/h2\u003e\n \u003cp\u003eThe semantic similarity Ss is calculated using a combination of techniques:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Object Set Similarity: Jaccard similarity between the sets of detected objects in the two images.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Semantic Embedding Similarity: Cosine similarity between the object-based semantic embeddings.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Graph Similarity: If using the graph-based representation, we employ graph matching techniques or graph kernels to compute similarity between the object relationship graphs.\u003c/p\u003e\n \u003c/span\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec25\" class=\"Section2\"\u003e\n \u003ch2\u003e3.7 Decision Module\u003c/h2\u003e\n \u003cp\u003eThe decision module determines whether two images are duplicates or near-duplicates based on the computed similarity score. We employ a threshold-based approach, where image pairs with a similarity score above a certain threshold are classified as duplicates. The threshold is determined through empirical analysis on a validation dataset, balancing precision and recall.\u003c/p\u003e\n \u003cdiv id=\"Sec26\" class=\"Section3\"\u003e\n \u003ch2\u003e3.7.1 Adaptive Thresholding\u003c/h2\u003e\n \u003cp\u003eTo account for variations in different image domains and use cases, we implement an adaptive thresholding mechanism. This approach adjusts the decision threshold based on the characteristics of the input images and the specific requirements of the application (e.g., prioritizing precision over recall or vice versa).\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec27\" class=\"Section3\"\u003e\n \u003ch2\u003e3.7.2 Confidence Estimation\u003c/h2\u003e\n \u003cp\u003eIn addition to the binary duplicate/non-duplicate decision, our framework provides a confidence estimate for each classification. This confidence score is derived from the distance between the computed similarity score and the decision threshold, allowing for more nuanced decision-making in borderline cases.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec28\" class=\"Section2\"\u003e\n \u003ch2\u003e3.8 Training and Optimization\u003c/h2\u003e\n \u003cp\u003eThe entire framework is trained end-to-end using a large dataset of labeled image pairs (duplicate and non-duplicate). We employ a multi-task learning approach, optimizing for both visual and semantic similarity simultaneously. The loss function combines:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Contrastive loss for visual similarity learning\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Cross-entropy loss for object recognition\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Graph-based loss for semantic relationship learning\u003c/p\u003e\n \u003c/span\u003e\n \u003cp\u003eOptimization is performed using stochastic gradient descent with momentum or adaptive learning rate methods such as Adam [39].\u003c/p\u003e\n\u003c/div\u003e"},{"header":"4. Experimental Setup and Evaluation Methodology","content":"\u003cp\u003eTo evaluate the effectiveness of our proposed semantic-aware image deduplication framework, we conducted a comprehensive set of experiments on multiple datasets. This section details our experimental setup, including datasets, evaluation metrics, and baseline methods for comparison.\u003c/p\u003e\n\u003cdiv id=\"Sec30\" class=\"Section2\"\u003e\n \u003ch2\u003e4.1 Datasets\u003c/h2\u003e\n \u003cp\u003eWe utilized the following datasets in our experiments:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. MNIST-Duplicate: A synthetic dataset created by applying random transformations (rotation, scaling, translation) to MNIST digits, resulting in 100,000 image pairs (50,000 duplicate pairs and 50,000 non-duplicate pairs).\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Oxford Buildings Dataset: A real-world dataset containing 5,062 images of Oxford landmarks, with ground truth information on duplicate and near-duplicate images [40].\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Web Image Dataset (WID): A large-scale dataset of 1\u0026nbsp;million web images, with known duplicate and near-duplicate pairs, collected from various online sources [41].\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e4. COCO-Duplicate: A custom dataset created from the COCO dataset [36] by applying various transformations and modifications to generate duplicate and near-duplicate pairs, focusing on images with multiple objects.\u003c/p\u003e\n \u003c/span\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eOverview of datasets used in the experiments\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"6\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDataset\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eImages\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDuplicate Pairs\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNon-Duplicate Pairs\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eObject Types\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTransformation Types\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMNIST-Duplicate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e200,000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e50,000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e50,000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDigits\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRotation, Scaling, Translation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOxford Buildings\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e5,062\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2,531\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2,531\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBuildings, Landmarks\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eViewpoint changes, Lighting variations\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWeb Image Dataset (WID)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1,000,000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e100,000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e900,000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVarious\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMultiple (natural variations)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCOCO-Duplicate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e200,000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e50,000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e50,000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMultiple object categories\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCropping, Color adjustments, Object addition/removal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec31\" class=\"Section2\"\u003e\n \u003ch2\u003e4.2 Evaluation Metrics\u003c/h2\u003e\n \u003cp\u003eTo assess the performance of our semantic-aware image deduplication framework, we employed the following evaluation metrics:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Precision: The ratio of correctly identified duplicate pairs to the total number of pairs identified as duplicates.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Recall: The ratio of correctly identified duplicate pairs to the total number of actual duplicate pairs in the dataset.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. F1-score: The harmonic mean of precision and recall, providing a balanced measure of the model\u0026apos;s performance.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e4. Mean Average Precision (mAP): A metric that considers the ranking of duplicate images, particularly useful for evaluating near-duplicate detection.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e5. Receiver Operating Characteristic (ROC) curve: A plot of the true positive rate against the false positive rate at various threshold settings.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e6. Area Under the ROC Curve (AUC): A single scalar value representing the overall performance across all possible thresholds.\u003c/p\u003e\n \u003c/span\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec32\" class=\"Section2\"\u003e\n \u003ch2\u003e4.3 Baseline Methods\u003c/h2\u003e\n \u003cp\u003eWe compared our semantic-aware framework against the following baseline methods:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Perceptual Hashing (pHash) [6]: A classic image hashing technique based on the discrete cosine transform.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. SIFT\u0026thinsp;+\u0026thinsp;BoVW: Scale-Invariant Feature Transform descriptors with Bag of Visual Words representation [10].\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Deep Siamese Network: A convolutional neural network trained with a contrastive loss function [34].\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e4. DeepRank: A deep learning-based approach for image similarity ranking [42].\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e5. DupNet: A recent deep learning method specifically designed for image deduplication [43].\u003c/p\u003e\n \u003c/span\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec33\" class=\"Section2\"\u003e\n \u003ch2\u003e4.4 Implementation Details\u003c/h2\u003e\n \u003cp\u003eOur semantic-aware image deduplication framework was implemented using PyTorch 1.8.0. The visual feature extraction module utilized a ResNet-50 architecture pre-trained on ImageNet, while the object recognition module employed a Faster R-CNN model pre-trained on COCO. The semantic feature extraction module used a Graph Convolutional Network with 3 layers.\u003c/p\u003e\n \u003cp\u003eTraining was performed on 4 NVIDIA Tesla V100 GPUs with 32GB memory each. We used the Adam optimizer with an initial learning rate of 1e-4 and a batch size of 64 image pairs. The model was trained for 50 epochs, with learning rate decay by a factor of 0.1 every 20 epochs.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec34\" class=\"Section2\"\u003e\n \u003ch2\u003e4.5 Experimental Protocol\u003c/h2\u003e\n \u003cp\u003eWe conducted our experiments following this protocol:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Data Preparation: Each dataset was split into training (60%), validation (20%), and test (20%) sets. Care was taken to ensure that duplicate pairs were not split across sets.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Model Training: The semantic-aware framework and baseline models were trained on the training set, with hyperparameters tuned using the validation set.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Threshold Selection: For each method, we determined the optimal decision threshold using the validation set, maximizing the F1-score.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e4. Performance Evaluation: We evaluated all methods on the test set using the selected thresholds and computed the metrics described in Section 4.2.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e5. Ablation Studies: We conducted ablation studies to analyze the impact of different components of our framework, including:\u003c/p\u003e\n \u003c/span\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003e○ Visual features only vs. Semantic features only vs. Combined features\u003c/p\u003e\n \u003cp\u003e○ Impact of different object recognition models\u003c/p\u003e\n \u003cp\u003e○ Contribution of graph-based semantic representation\u003c/p\u003e\n \u003c/div\u003e\n \u003cp\u003e6. Scalability Analysis: We assessed the computational requirements and scalability of our approach by measuring processing time and memory usage on datasets of increasing size.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec35\" class=\"Section2\"\u003e\n \u003ch2\u003e4.6 Statistical Analysis\u003c/h2\u003e\n \u003cp\u003eTo ensure the statistical significance of our results, we performed the following analyses:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Confidence Intervals: We computed 95% confidence intervals for all reported metrics using bootstrap resampling with 1000 iterations.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Paired t-tests: We conducted paired t-tests to compare the performance of our method against each baseline, with a significance level of \u0026alpha;\u0026thinsp;=\u0026thinsp;0.05.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. McNemar\u0026apos;s Test: For binary classification decisions (duplicate vs. non-duplicate), we used McNemar\u0026apos;s test to assess the statistical significance of improvements over baseline methods.\u003c/p\u003e\n \u003c/span\u003e\n\u003c/div\u003e"},{"header":"5. Results and Discussion","content":"\u003cp\u003eIn this section, we present the results of our experimental evaluation and provide a detailed analysis of the performance of our semantic-aware image deduplication framework compared to baseline methods.\u003c/p\u003e\n\u003cdiv id=\"Sec37\" class=\"Section2\"\u003e\n \u003ch2\u003e5.1 Overall Performance Comparison\u003c/h2\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e shows the performance of our proposed method and baseline approaches across all datasets, reporting precision, recall, F1-score, and MAP.\u003c/p\u003e\n \u003cp\u003eOur semantic-aware framework consistently outperforms all baseline methods across all datasets, achieving significant improvements in both F1-score and mAP. The performance gain is particularly notable on the COCO-Duplicate dataset, which contains images with multiple objects and complex semantic relationships.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec38\" class=\"Section2\"\u003e\n \u003ch2\u003e5.2 ROC Curve Analysis\u003c/h2\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e presents the ROC curves for all methods on the Web Image Dataset (WID), illustrating the trade-off between true positive rate and false positive rate across different decision thresholds.\u003c/p\u003e\n \u003cp\u003eThe ROC curve analysis reveals that our semantic-aware approach maintains a higher true positive rate across all false positive rate levels, indicating superior performance in distinguishing between duplicate and non-duplicate images. The Area under the Curve (AUC) values are reported in Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv align=\"char\" class=\"colspec\"\u003e\u003cbr\u003e\u003c/div\u003e\u0026nbsp;\u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eAUC values for different methods on the Web Image Dataset\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"2\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMethod\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003epHash\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.867\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSIFT\u0026thinsp;+\u0026thinsp;BoVW\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.901\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDeep Siamese\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.953\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDeepRank\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.961\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDupNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.968\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOurs (Semantic-Aware)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.984\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec39\" class=\"Section2\"\u003e\n \u003ch2\u003e5.3 Ablation Studies\u003c/h2\u003e\n \u003cp\u003eTo understand the contribution of different components in our framework, we conducted ablation studies. Table \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e presents the results of these studies on the COCO-Duplicate dataset.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eAblation study results on COCO-Duplicate dataset\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMethod\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003emAP\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVisual features only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.891\u0026thinsp;\u0026plusmn;\u0026thinsp;0.005\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.878\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSemantic features only\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.903\u0026thinsp;\u0026plusmn;\u0026thinsp;0.004\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.891\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCombined (no graph)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.918\u0026thinsp;\u0026plusmn;\u0026thinsp;0.003\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.909\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFull model (with graph)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.929\u0026thinsp;\u0026plusmn;\u0026thinsp;0.003\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.921\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 ablation studies demonstrate that both visual and semantic features contribute significantly to the overall performance. The integration of graph-based semantic representation further enhances the accuracy of our framework.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec40\" class=\"Section2\"\u003e\n \u003ch2\u003e5.4 Impact of Object Recognition Models\u003c/h2\u003e\n \u003cp\u003eWe experimented with different object recognition models to assess their impact on the framework\u0026apos;s performance. Table 5 shows the results using various object detection architectures on the COCO-Duplicate dataset.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab4\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003ePerformance with different object detection models on COCO-Duplicate dataset.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eObject Detection Model\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eF1-score\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003emAP\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFaster R-CNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.929\u0026thinsp;\u0026plusmn;\u0026thinsp;0.003\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.921\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMask R-CNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.932\u0026thinsp;\u0026plusmn;\u0026thinsp;0.003\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.924\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYOLO v4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.925\u0026thinsp;\u0026plusmn;\u0026thinsp;0.004\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.917\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEfficientDet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.930\u0026thinsp;\u0026plusmn;\u0026thinsp;0.003\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.922\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\u003eWhile all object detection models provide strong performance, Mask R-CNN shows a slight edge, likely due to its ability to provide more detailed object segmentation information.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec41\" class=\"Section2\"\u003e\n \u003ch2\u003e5.5 Scalability Analysis\u003c/h2\u003e\n \u003cp\u003eTo assess the scalability of our approach, we measured the processing time and memory usage for deduplication tasks of increasing dataset sizes. Figure \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e illustrates the relationship between dataset size and computational resources required.\u003c/p\u003e\n \u003cp\u003eOur semantic-aware framework shows a linear increase in processing time and memory usage as the dataset size grows, indicating good scalability for large-scale deduplication tasks. However, it does require more computational resources compared to simpler methods like perceptual hashing.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec42\" class=\"Section2\"\u003e\n \u003ch2\u003e5.6 Qualitative Analysis\u003c/h2\u003e\n \u003cp\u003eOur semantic-aware approach demonstrates superior performance in handling cases where images are semantically similar but visually distinct, such as different views of the same landmark or products with varying backgrounds.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec43\" class=\"Section2\"\u003e\n \u003ch2\u003e5.7 Discussion\u003c/h2\u003e\n \u003cp\u003eThe experimental results clearly demonstrate the effectiveness of our semantic-aware image deduplication framework. Key findings include:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Consistent Performance Improvement: Our approach consistently outperforms existing methods across all datasets, with statistically significant improvements in F1-score and mAP.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Robustness to Visual Variations: The semantic-aware framework shows particular strength in handling cases where duplicate images have significant visual differences but share semantic content.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Complementary Features: The ablation studies confirm that both visual and semantic features contribute to the overall performance, with the graph-based semantic representation providing additional gains.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e4. Scalability Trade-offs: While our method requires more computational resources than simpler approaches, it scales linearly with dataset size, making it feasible for large-scale applications.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e5. Flexibility: The framework\u0026apos;s performance can be further optimized by choosing appropriate object detection models for specific use cases or domains.\u003c/p\u003e\n \u003c/span\u003e\n \u003cp\u003eDespite these promising results, there are some limitations and areas for future work:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Computational Overhead: The inclusion of object recognition and graph-based processing increases computational requirements, which may be a concern for extremely large-scale or real-time applications.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Domain Specificity: While our approach generalizes well across the tested datasets, further investigation is needed to assess its performance on highly specialized image collections (e.g., medical imaging, satellite imagery).\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Fine-grained Similarity: Future work could explore more nuanced similarity metrics that consider the relative importance of different objects or semantic relationships within images.\u003c/p\u003e\n \u003c/span\u003e\n \u003cp\u003eIn conclusion, our semantic-aware image deduplication framework represents a significant advancement in the field, effectively leveraging both visual and semantic information to achieve state-of-the-art performance in identifying duplicate and near-duplicate images.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"6. Conclusion and future work","content":"\u003cp\u003eThis paper presented a novel semantic-aware image deduplication framework that leverages object recognition techniques to enhance the accuracy of duplicate and near-duplicate image detection. Our approach integrates visual feature analysis with semantic information derived from state-of-the-art object detection models, demonstrating significant improvements over existing methods across diverse datasets.\u003c/p\u003e\n\u003cp\u003eKey contributions of this work include:\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e1. A comprehensive framework that combines visual and semantic features for image deduplication.\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e2. A novel graph-based semantic representation that captures complex relationships between objects in images.\u003c/p\u003e\n\u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e3. Extensive experimental evaluation demonstrating the effectiveness and scalability of the proposed approach.\u003c/p\u003e\n\u003c/span\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eThe results of our experiments highlight the potential of semantic-aware techniques in addressing the limitations of traditional image deduplication methods, particularly in scenarios involving visually distinct but semantically similar images.\u003c/p\u003e\n\u003cp\u003eFuture research directions include:\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e1. Exploring more advanced semantic representations: Investigating the use of scene graphs or visual relationship detection to capture even more nuanced semantic information.\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e2. Incorporating temporal information: Extending the framework to handle video deduplication by considering temporal relationships between frames.\u003c/p\u003e\n\u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e3. Adaptive feature weighting: Developing mechanisms to dynamically adjust the importance of visual and semantic features based on the characteristics of the input images.\u003c/p\u003e\n\u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e4. Privacy-preserving deduplication: Investigating techniques to perform semantic-aware deduplication while preserving user privacy, especially for sensitive applications.\u003c/p\u003e\n\u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e5. Cross-modal deduplication: Extending the framework to identify duplicates across different modalities, such as images and text descriptions.\u003c/p\u003e\n\u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e6. Unsupervised and self-supervised learning: Exploring ways to reduce the reliance on large labeled datasets for training the deduplication model.\u003c/p\u003e\n\u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e7. Edge deployment: Optimizing the framework for deployment on edge devices with limited computational resources.\u003c/p\u003e\n\u003c/span\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eIn conclusion, semantic-aware image deduplication represents a promising direction for improving the accuracy and robustness of duplicate detection systems. As the volume and complexity of digital images continue to grow, such advanced techniques will become increasingly important for efficient data management and content curation across various domains.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eConflicts of interest\u003c/h2\u003e \u003cp\u003eThe authors declare that they have no conflict of interests.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eEthics, Consent to Participate, and Consent to Publish declarations\u003c/h2\u003e \u003cp\u003eNot applicable.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThis research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eR.S. (Rahul Shah) conceptualized the study and designed the research framework. R.S. and A.K.S. (Dr. Ashok Kumar Shrivastava) developed the methodology and conducted the experiments. A.K.S. contributed to data collection and preprocessing. R.S. performed the analysis and interpretation of results. R.S. wrote the main manuscript text, and A.K.S. reviewed and provided critical revisions. Both authors approved the final version of the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgment\u003c/h2\u003e \u003cp\u003eWe would like to express our sincere gratitude to everyone who contributed to this research. Special thanks to our colleagues for their insightful feedback and support throughout the research. We are also grateful to the institutions and organizations that provided the resources and datasets necessary for this study. Finally, we thank our families and friends\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data supporting the findings of this study are available from the corresponding author upon reasonable request. Due to certain restrictions, if applicable, such as privacy, confidentiality, or proprietary concerns, the dataset is not publicly accessible.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWang, X., et al. A survey on visual content-based video indexing and retrieval. *IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews)*. 2014; 41(6): 797\u0026ndash;819.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChum, O., et al. (2008). Near duplicate image detection: min-hash and tf-idf weighting. In BMVC (Vol. 810, pp. 812\u0026ndash;815).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZauner, C. (2010). Implementation and benchmarking of perceptual image hash functions. Master's thesis, University of Applied Sciences Hagenberg, Austria.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKe, Y., et al. (2004). Efficient near-duplicate detection and sub-image retrieval. In Proceedings of the 12th annual ACM international conference on Multimedia (pp. 869\u0026ndash;876).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang, J., et al. (2011). Image indexing using color correlograms. In 2011 IEEE Conference on Computer Vision and Pattern Recognition (pp. 762\u0026ndash;768).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMonga, V., \u0026amp; Evans, B. L. (2006). Perceptual image hashing via feature points: performance evaluation and tradeoffs. IEEE transactions on Image Processing, 15(11), 3452\u0026ndash;3465.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, B., et al. (2006). A comparative study on robust hash functions for digital audio. In International Workshop on Digital Watermarking (pp. 354\u0026ndash;365). Springer, Berlin, Heidelberg.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ\u0026eacute;gou, H., et al. (2008). Hamming embedding and weak geometric consistency for large scale image search. In European conference on computer vision (pp. 304\u0026ndash;317). Springer, Berlin, Heidelberg.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSmeulders, A. W., et al. (2000). Content-based image retrieval at the end of the early years. IEEE Transactions on pattern analysis and machine intelligence, 22(12), 1349\u0026ndash;1380.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLowe, D. G. (2004). Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2), 91\u0026ndash;110.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBay, H., et al. (2008). Speeded-up robust features (SURF). Computer vision and image understanding, 110(3), 346\u0026ndash;359.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, Y., et al. (2007). A survey of content-based image retrieval with high-level semantics. Pattern recognition, 40(1), 262\u0026ndash;282.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeCun, Y., et al. (2015). Deep learning. nature, 521(7553), 436\u0026ndash;444.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKrizhevsky, A., et al. (2012). Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 1097\u0026ndash;1105.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGirshick, R., et al. (2014). Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 580\u0026ndash;587).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGirshick, R. (2015). Fast r-cnn. In Proceedings of the IEEE international conference on computer vision (pp. 1440\u0026ndash;1448).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRedmon, J., et al. (2016). You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 779\u0026ndash;788).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe, K., et al. (2017). Mask r-cnn. In Proceedings of the IEEE international conference on computer vision (pp. 2961\u0026ndash;2969).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTan, M., et al. (2020). Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 10781\u0026ndash;10790).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGordo, A., et al. (2016). Deep image retrieval: Learning global representations for image search. In European conference on computer vision (pp. 241\u0026ndash;257). Springer, Cham.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSimonyan, K., \u0026amp; Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe, K., et al. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770\u0026ndash;778).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBabenko, A., \u0026amp; Lempitsky, V. (2015). Aggregating local deep features for image retrieval. In Proceedings of the IEEE international conference on computer vision (pp. 1269\u0026ndash;1277).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, H., et al. (2018). Visual semantic reasoning for image-text matching. In Proceedings of the IEEE International Conference on Computer Vision (pp. 4654\u0026ndash;4662).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLong, J., et al. (2015). Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 3431\u0026ndash;3440).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, L. C., et al. (2017). Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4), 834\u0026ndash;848.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao, H., et al. (2017). Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2881\u0026ndash;2890).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu, C., et al. (2016). Visual relationship detection with language priors. In European conference on computer vision (pp. 852\u0026ndash;869). Springer, Cham.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, H., et al. (2017). Visual translation embedding network for visual relation detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 5532\u0026ndash;5540).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZellers, R., et al. (2018). Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 5831\u0026ndash;5840).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZheng, L., et al. (2017). SIFT meets CNN: A decade survey of instance retrieval. IEEE transactions on pattern analysis and machine intelligence, 40(5), 1224\u0026ndash;1244.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, J., et al. (2017). Stacked attention networks for image question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 21\u0026ndash;29).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTan, M., \u0026amp; Le, Q. (2019). Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning (pp. 6105\u0026ndash;6114). PMLR.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKoch, G., et al. (2015). Siamese neural networks for one-shot image recognition. In ICML deep learning workshop (Vol. 2).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRen, S., et al. (2015). Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 91\u0026ndash;99.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin, T. Y., et al. (2014). Microsoft coco: Common objects in context. In European conference on computer vision (pp. 740\u0026ndash;755). Springer, Cham.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLazebnik, S., et al. (2006). Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06) (Vol. 2, pp. 2169\u0026ndash;2178). IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKipf, T. N., \u0026amp; Welling, M. (2016). Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKingma, D. P., \u0026amp; Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePhilbin, J., et al. (2007). Object retrieval with large vocabularies and fast spatial matching. In 2007 IEEE conference on computer vision and pattern recognition (pp. 1\u0026ndash;8). IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, X. J., et al. (2013). Duplicate discovery on 2 billion internet images. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 429\u0026ndash;436).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"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":"Image Deduplication, Object Recognition, Semantic Similarity, Deep Learnings, Computer Vision, Content-Based Image Retrieval","lastPublishedDoi":"10.21203/rs.3.rs-6396148/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6396148/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eImage deduplication is a critical task in managing large-scale image databases, with applications ranging from content moderation to efficient storage solutions. Traditional methods often rely on pixel-level comparisons or perceptual hashing techniques, which may fail to capture semantic similarities between images. This paper proposes a novel approach to image deduplication that incorporates semantic information derived from object recognition algorithms. By leveraging state-of-the-art deep learning models for object detection and classification, our method achieves higher accuracy in identifying duplicate and near-duplicate images, even in cases where visual features may differ significantly. We present a comprehensive evaluation of our semantic-aware deduplication system on diverse datasets, demonstrating its superior performance compared to existing methods. Our findings suggest that integrating semantic understanding into image deduplication processes can significantly enhance accuracy and robustness, paving the way for more efficient and intelligent image management systems.\u003c/p\u003e","manuscriptTitle":"Semantic-Aware Image Deduplication: Leveraging Object Recognition for Enhanced Accuracy","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-30 14:48:44","doi":"10.21203/rs.3.rs-6396148/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":"08400df4-9ab4-4251-a472-fd283c2fd20f","owner":[],"postedDate":"May 30th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-08-21T17:50:54+00:00","versionOfRecord":{"articleIdentity":"rs-6396148","link":"https://doi.org/10.22266/ijies2025.0831.09","journal":{"identity":"international-journal-of-intelligent-engineering-and-systems","isVorOnly":true,"title":"International Journal of Intelligent Engineering and Systems"},"publishedOn":"2025-08-02 00:00:00","publishedOnDateReadable":"August 2nd, 2025"},"versionCreatedAt":"2025-05-30 14:48:44","video":"","vorDoi":"10.22266/ijies2025.0831.09","vorDoiUrl":"https://doi.org/10.22266/ijies2025.0831.09","workflowStages":[]},"version":"v1","identity":"rs-6396148","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6396148","identity":"rs-6396148","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}