Classification and Diagnosis of Defects in Steel Surfaces Using Deep Convolutional Neural Networks

Classification and Diagnosis of Defects in Steel Surfaces Using Deep Convolutional Neural Networks | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 6 February 2025 V1 Latest version Share on Classification and Diagnosis of Defects in Steel Surfaces Using Deep Convolutional Neural Networks Author : Y. Shobha 0009-0007-5527-194X [email protected] Authors Info & Affiliations https://doi.org/10.22541/au.173882390.08755910/v1 375 views 129 downloads Contents Abstract Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Detecting surface defects in steel manufacturing is crucial for product quality and production efficiency. However, real-time quality control faces challenges in automation and reliability. Surface flaws in steel strips vary in complexity, requiring robust defect detection algorithms with high generalization performance.For the purpose of addressing this issue, we visited JSW Steel Ltd, Vijayanagara, Ballari, Karnataka, India, and took pictures of any defects we found. Using deep convolutional neural networks (CNNs), we provide a new method for steel defect identification. Images of surface defects in steel were used to train and evaluate the deep CNN. This nine-layer CNN model solves the problem of finding flaws in steel strips as a whole. Implementation of data augmentation techniques to avoid overfitting. The performance of the proposed model was assessed using a dataset that included three classes of flaws and one class that was free of flaws. On the validation set, the model achieved an impressive accuracy of 93.27 percent. The experimental results demonstrate the deep CNN model’s higher performance for both intra and inter-class fault detection. So, the proposed deep CNN model gives a precise and real-time way to find surface flaws in steel strip production lines, leading to higher-quality steel strips overall. jabbrv-ltwa-all.ldf jabbrv-ltwa-en.ldf Classification and Diagnosis of Defects in Steel Surfaces Using Deep Convolutional Neural Networks Shobha Y Ballari Business College, Vijayanagara Sri Krishnadevaraya University, Kishkinda University Ballari, Karnataka-India. Abstract Detecting surface defects in steel manufacturing is crucial for product quality and production efficiency. However, real-time quality control faces challenges in automation and reliability. Surface flaws in steel strips vary in complexity, requiring robust defect detection algorithms with high generalization performance.For the purpose of addressing this issue, we visited JSW Steel Ltd, Vijayanagara, Ballari, Karnataka, India, and took pictures of any defects we found. Using deep convolutional neural networks (CNNs), we provide a new method for steel defect identification. Images of surface defects in steel were used to train and evaluate the deep CNN. This nine-layer CNN model solves the problem of finding flaws in steel strips as a whole. Implementation of data augmentation techniques to avoid overfitting. The performance of the proposed model was assessed using a dataset that included three classes of flaws and one class that was free of flaws. On the validation set, the model achieved an impressive accuracy of 93.27 percent. The experimental results demonstrate the deep CNN model’s higher performance for both intra and inter-class fault detection. So, the proposed deep CNN model gives a precise and real-time way to find surface flaws in steel strip production lines, leading to higher-quality steel strips overall. Keywords: DeepCNN,steel surface defect, intra class and inter-class,Rollmark, Scale, Scratches and Normal. Introduction Steel, a highly adaptable metallic substance, is widely utilized in a multitude of domains that encompass our daily existence. This renders it an optimal selection for an array of sectors, including civil engineering infrastructure, aerospace, shipbuilding, automotive, machinery manufacturing, and household tool production.NirbharNeogi et al.[1] propose a classification scheme for steel surfaces, dividing them into two primary categories: flat products and long products. The classification of flat products includes several subcategories, specifically slab/billet, steel plate, and hot/cold-rolled strips. Hot-rolled steel strips refer to flat steel products that are manufactured through the process of hot-rolling a steel slab or billet. The aforementioned strips possess elongated and slender dimensions, exhibiting consistent measurements in terms of both thickness and width. The hot rolling process is a multi-step operation that begins with melting steel and casting it into a continuous billet. As described by Nong Jin et al. [2], the billet goes through a series of passes in a hot rolling mill, starting with a roughing pass and ending with a finishing pass. This transformation results in a rod, which is then employed in several subsequent operations, such as forging.However, a number of variables, such as equipment deterioration and changes in the production process, might affect the output of hot-rolled strips. As a result, these elements might cause manufacturing losses in the form of surface flaws on the strips. The prevalence of surface defects is a key problem in the steel industry since empirical research shows that the majority of these problems (85%) develop on the surfaces of steel workplaces in real-world situations. Slivers, scales, scratches, scabs, pits, and roll marks are just some of the prevalent surface flaws found on hot-rolled items. Both the items’ aesthetic value and their functionality might be negatively affected by these flaws. Munoz-Escalona et al.[3] employ the manual inspection technique to find surface imperfections in steel strips. This method, however, was known for being difficult to implement and unstable over time. The process’s high workload and high volume of output were cited as the primary causes of the problems. According to studies conducted byJinrui Gan et al. [4], ASI systems are quicker and more reliable than manual inspections. The introduction of automated systems has revolutionized the inspection process, resulting in greater accuracy and efficiency. As manual inspection is employed on only a small percentage of the manufacturing floor (usually less than 0.05 percent), current research efforts are concentrated on improving automated surface inspection (ASI) systems. ASI for steel typically consists of two main steps: feature extraction and fault categorization. Local binary pattern by Matti Niskanen et al. [5], SIFT and Voting Strategy by Suvdaa et al. [6], wavelet and Support Vector Machine by Santanu Ghorai et al. [7] are just a few of the techniques used in feature extraction to give complete representations of defect images. Machine Learning (ML) is a tried-and-true method for defect detection that has been widely employed in previous studies. In this method, human-created feature extractors are used to gather relevant data. Following this, other classifiers are used to learn these features, such as the K-nearest neighbors by FlorentDupont et al. [8] and the random forest by Bae-Keun Kwon et al. [9]. When it comes to classification and prediction tasks, previous work by Prasad et al. [10, 11] has demonstrated the use of machine learning (ML) and artificial neural networks (ANN). In addition, Vaidya et al. [12] showed how to extract features using Artificial Neural Networks (ANN) and Principal Component Analysis (PCA). Traditional methods have been shown to be effective in defect identification, but they have drawbacks in terms of limited resilience and real-time speed. For a variety of tasks, including data pre-processing, feature extraction, feature reduction, and classifier selection, these algorithms rely heavily on the knowledge and experience of humans. Therefore, people often struggle to correctly identify actual fault borders in circumstances with noisy or textured backdrops. This shortcoming reduces their use for satisfying the needs of online surface flaw identification. Alex Krizhevsky et al.[13], Karen Simonyan and Andrew Zisserman[14],Sergey Zagoruykoand Nikos Komodakis[15] have all done research showing that the Convolutional Neural Network (CNN) is the best deep learning model, being very good at a wide range of recognition tasks and giving results that are at the cutting edge. Using maxpooling convolutional neural networks, Jonathan Masci et al.[16] presented a solution for surface defect identification with a 93% accuracy rate. Daniel Weimer et al. [17] designed Convolutional Neural Network in industrial inspection In addition, Long Wen et al.[18] have shown its practical value in practice.Gao Yiping et al.[19] presented a unique technique that incorporates a convolutional neural network (CNN) in the field of semi-supervised learning. Using this method has the dual advantage of requiring a smaller number of labelled samples while also making use of the unrealized potential of unlabelled data for training. Further, Prasad et al.’s [20] study focused on the use of Convolutional Neural Networks (CNNs) for the objective of robotic feature extraction. In contrast to more typical machine learning approaches, convolutional neural networks (CNNs) may learn features on their own. With this improvement, CNN could skip feature extraction and get straight into processing pictures. As a result, researchers and developers have focused on expanding and refining CNN-based approaches to issue detection. In this article, we’ll take a look at some of the most recent developments and groundbreaking discoveries in the science of inspecting steel for defects. The final purpose of this proposed study is to provide a standardized approach to accurately detecting surface flaws on steel strips. In this research, we used deep convolutional neural networks (CNNs). A clipped defect image from a rolling mill at JSW Steel Ltd. is used as input to activate the method. The network will then make an educated guess as to how to label the problem. Unlike competing approaches, ours does not necessitate the development of unique feature extraction procedures thanks to the use of Convolutional Neural Network (CNN) models that have been trained on raw fault images. This eliminates the need for a feature extraction phase, as the network learns about defect characteristics as part of the training process. To combat the issue of overfitting brought on by highly networked data, we also employ a wide range of data augmentation strategies. This practice ensures that our model exhibits effective generalization capabilities when presented with previously unseen examples. Our methodology addresses both inter-class and intra-class defect challenges. Intra-class defects demonstrate notable disparities in their visual characteristics, resulting in a considerable degree of visual diversity within a given class. Conversely, inter-class defects demonstrate comparable attributes, leading to the presence of common visual elements among various defect classifications. Intra-class defects exhibit a wide dispersion of features, indicative of their diverse appearances, whereas inter-class defects display notable similarities in their features. This paper makes sincere efforts to address the following issues 1. In the steel manufacturing industries, how important is it to detect surface flaws, and how do these defects affect product quality and production efficiency? 2. Why aren’t automated and reliable real-time diagnostics available for steel manufacturing yet, and what are the limits of the ones that are? 3. The proposed deep Convolutional Neural Network (CNN) model’s creation and assessment rely on a dataset comprised of defect photos gathered from a steel factory; what is the importance of this dataset? 4. When it comes to finding surface flaws on steel strips, how can data augmentation approaches help prevent overfitting, and how does the deep CNN model with nine convolutional layers give an end-to-end solution? 5. To what extent does the suggested model accurately detect and categorize the various types of flaws contained in the dataset? 6. The experimental results show that the proposed DCNN model can effectively recognize defects both within and across classes. The rest of the information is laid down as follows: The specifics of the planned DCNN are laid out in Section 2. In Section 3, we detail our experience putting the suggested method through its paces on a carefully selected and curated dataset. The final part of the article presents the study’s conclusion and suggested subsequent measures. Materials and Methods jabbrv-ltwa-all.ldf jabbrv-ltwa-en.ldf 2.1.Dataset Description This research relied on a dataset gathered from JSW Steel Ltd, Ballari, Karnataka, India. There are 30,373 photos in all, split between defect categories (scratch, rollmark, and scale) and normal (no defect) categories depicting steel strips. There are 25,702 photos in the training folder and 4,671 in the testing folder. The training folder contains a total of 6517 photos, with 6301 located in the scratch folder, 6332 in the roll mark folder, 6552 in the scale folder, and 6517 in the normal folder. There are a total of 1065 photographs in the testing folder’s scratch folder, 1711 images in the rollmark folder, 1276 images in the scale folder, and 619 images in the normal folder. Scratches, rollmarks, scales, and normal pictures are the four types of common surface flaws seen in Figure 1. Initially, the picture was 720 pixels wide by 128 pixels high. The original photos are down-sampled in this work, with the sampled images set at 150x150 pixels to conserve computational resources for later portions. The Dataset also comprises of both inter and intra-class images. 2.2.Data Augmentation When it comes to image identification, deep learning relies on preprocessing pictures to extract useful characteristics. Data augmentation is a common practice in the realm of preprocessing. To enhance data, one modifies it without changing its essential nature. This method speeds up the preprocessing phase, which is very helpful when working with massive amounts of data. The following enhancement strategies were used in this investigation: 1. Data augmentation’s rescaling (1. /255) process divides image pixel values by 255 to place them in the [0, 1] range. This normalization aids the training of neural networks for computer vision applications by enhancing their convergence and stability. 2. Shear range of 0.1: Shearing is a data augmentation technique that slants the image’s shape along an axis. With a shear range of 0.1, the image can be deformed by up to 10% of its width or height. This enhances the model’s robustness to varying perspectives and angles 3. Zoom range of 0.1: With a zoom range of 0.1, the image can be zoomed in or out by up to 10%. This augmentation improves the model’s ability to recognize objects at different scales, enhancing its versatility in handling objects of varying sizes. 4. Brightness range between 0.6 and 1.0: A brightness range of 0.6 to 1.0 allows the image’s brightness to be adjusted by up to 40% higher or 60% lower than its original value. This reduces the model’s sensitivity to changes in lighting conditions. 5. Horizontal flip:Horizontal flipping is a data augmentation technique that mirrors the image horizontally, making the left side become the right side and vice versa, useful for orientation-agnostic tasks like object recognition 6. Vertical flip:Vertical flipping is a data augmentation technique that mirrors the image vertically, swapping the top and bottom parts. Like horizontal flipping, it’s beneficial for orientation-agnostic tasks, where object classification doesn’t depend on orientation. Figure1.Categorical examples of surface defect images 2.3.Proposed Methodology The next parts will delve into the proposed model, which makes use of proposed deep CNN’s many layers.The steps for analysing photographs of steel surface defects are shown in Figure 2 as a block diagram. Images are processed, enhanced, trained, and evaluated throughout these phases.Using steel surface defect photos obtained from a steel mill, our suggested method makes use of Deep CNN architecture as the basis for fault identification. Training and hyperparameter tuning are also a part of this procedure. By fine-tuning the neural network’s biases and learning rate, an optimizer can help bring about a net decrease in loss and an increase in accuracy.In machine learning, the loss function quantifies an algorithm’s performance in terms of how well it fits the input data. The loss function learns to optimize itself over time, reducing prediction errors. The Adam optimizer is used in conjunction with a loss function based on categorical cross entropy to solve this problem. Figure2.Workflow diagram of proposed deep CNN model 2.4.Deep CNN Architecture Convolutional Neural Networks (CNNs) are multi-layer deep neural networks optimized for processing and analysing two-dimensional pictures. It was first proposed in the work of YanLeCun et al. [21], and it has since been shown to be useful for a variety of image processing and recognition applications. Both feature extraction and categorization are important components of the standard CNN architecture. As it progresses through the layers, the network learns to identify and characterize various spatial features in the input pictures. The collected characteristics are subsequently sent into the model’s classification phase, where probabilities are assigned. When a neural network is trained using the backpropagation technique, the model’s weights and biases are optimized. Convolutional layers, pooling layers, a fully connected layer, and a softmax layer are the building blocks of a CNN network. The CNN’s input layer is responsible for processing data when it is first fed into the network. When an image is read in, this layer stands in for the picture’s pixel matrix. The pixel matrix, or three-dimensional matrix, is fed into the neural network at the input layer, where it undergoes a series of transformations before arriving at the fully connected layer. One of the most important parts of a CNN is its convolution layer. From the preceding layer of the neural network, it gets a subset of the data. Although it only receives a subset of the data, the convolution layer analyses it extensively to derive more generalized characteristics. When compared to the convolution layer, the pooling layer’s function is rather different. The pooling layer shrinks the node matrix it gets from the layer above it. This size reduction aids in lowering overall neural network parameters and the number of nodes in the fully connected layer. CNN’s completely linked layer is found in the last two layers. Its main purpose is to return classifications for data that has been processed by the multilayer convolution and pooling layers. In a CNN, the convolution and pooling layers collaborate to automatically extract picture features and improve the information content, while the fully connected layer takes on the remaining responsibilities of data collection, classification, and summary. The softmax layer is used for classification tasks, and it does so by assigning probabilities to the data at hand.Convolutional neural networks are seen in their rudimentary form in Figure3. Figure3.Basic structure of Convolutional Neural Network. 2.5.Proposed Model Figure4 provides an overview of the architectural design of the proposed deep CNN model, while Table 1 presents detailed information about its constituent layers. The model encompasses multiple layers, including convolutional, maxpooling, and dropout layers, to process the input, which is a 150×150 pixel image. In the first block, two convolutional layers with 32 filters and a 3×3 kernel size operate on the input with a shape of 150×150×32, yielding the same output shape. The second block introduces two additional convolutional layers with the same specifications, resulting in an output shape of 75×75×32 from the previous block. Moving to the third block, it consists of two convolutional layers with 64 filters and a 3×3 kernel size. Given an input shape of37× 37×32, this block generates an output shape of 37×37×64. The fourth block follows a similar pattern, comprising two convolutional layers with 32 filters and a 3×3 kernel size. Utilizing the output from the previous block with a shape of 18×18×64 this block produces an output shape of 18×18×32. Finally, the fifth block consists of a single convolutional layer that takes the output from the previous block, which has a shape of 9×9×32, and generates an output shape of 9×9×64. Each convolutional layer is accompanied by a pooling layer and a dropout layer.ReLU activation is applied to introduce non-linearity and enhance the model’s learning and fitting capabilities. Maxpoolingis usually used in CNN due to the fast convergence and enhances model’s robustness as explained by Lan Boureauet al.[22].Pooling layers are employed to reduce the number of parameters; thereby accelerating the training process.Following the pooling layers, a dropout layer with a dropout rate of 0.01 is incorporated. Extensive research has demonstrated that dropout can be considered a form of data augmentation and effectively mitigate overfitting as reported according to the studies of Xavier Bouthillieret al.[23]. During the dropout phase, a random subset of neurons is deactivated based on the specified dropout rate, suspending their contribution to the activation of downstream neurons during the forward pass. Furthermore, weight updates are not applied to these deactivated neurons during the backward pass. Additionally, a flattening layer is included in the model’s architecture to convert the output from the preceding layer, which is a 4D tensor, into a 2D tensor with a shape of 5184. Lastly, an additional dense layer with 128 neurons and a softmax classifier with 4 nodes is incorporated to facilitate the classification task. Figure4.Schematic representation of the proposed deep CNN Model Table 1.Proposed Structure of the Model Input N/A (1,150,150) Convolution (3,3) (32,150,150) Convolution (3,3) (32,150,150) Max Pooling (2,2) (32,75,75) Convolution (3,3) (32,75,75) Convolution (3,3) (32,75,75) Max Pooling (2,2) (32,37,37) Convolution (3,3) (64,37,37) Convolution (3,3) (64,37,37) MaxPooling (2,2) (64,18,18) Convolution (3,3) (32,18,18) Convolution (3,3) (32,18,18) MaxPooling (2,2) (32,9,9) Convolution (3,3) (64,9,9) Flatten N/A (5120) Softmax N/A (4) Performance Evaluation jabbrv-ltwa-all.ldf jabbrv-ltwa-en.ldf 3.1.Hyperparameters The suggested deep CNN model depends on both accuracy and the loss function to achieve its full potential. As a smaller estimated loss indicates more efficiency, reducing the error rate is the major goal of the suggested approach. The authors here used categorical cross-entropy (CE), a well-liked loss function for multi-class classification problems. The Adam optimizer was used to help improve the proposed deep CNN model and decrease training loss. Adam is a method for optimizing a model’s parameters that uses adaptive gradient descent to quickly converge to a local minimum. Adam was selected above other optimization methods like SGD and RMSProp because of its simplicity, memory efficiency, and speed of learning. The values used for the hyperparameters are listed in Table 2, provided by the authors. The model’s LR is set to a modest value that is compatible with the other hyperparameters for optimal performance. The researchers found that Adam converged quickly and efficiently to optimal solutions. A batch size of 30 is used in our experimental investigation to facilitate data transfer throughout the network and efficient management of computational memory. The model is trained for a set number of iterations, or ”epochs,” in order to assess its performance over time. Performance of the proposed model is evaluated using Algorithm1.Top of FormTop of Form 1. Load the input images. 2. Apply data augmentation techniques. 3. Split the dataset into training and testing sets. 4. Initialize and construct the CNN model. 5. for each epoch in the specified number of epochs: for each batch in the specified batch size: Compute the predicted outputs ŷ by passing the batch of features through the model.Calculate the loss using the cross-entropy function with the true labels y.Optimize the model’s parameters based on the loss.Calculate the accuracy of the model on the current batch.Keep track of the best accuracy achieved during training. 6. Return the best accuracy achieved. Algorithm1: Algorithm for Performance Evaluation and Metric Analysis Table 2.Representation of Hyperparameter values 1 Optimizer Adam 2 Learning Rate 0.0001 3 Batch size 30 4 Epochs 10 5 Loss Function Categorical cross-entropy The results of a classification prediction can be tabulated and summarized using a confusion matrix. It allows for a contrast to be drawn between correct and faulty predictions, illuminating potential blind spots in the model. Results are broken down into four groups using the confusion matrix: true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). The performance of the model may be evaluated based on the criteria shown in Table 3. Table 3.Elements of Confusion Matrix jabbrv-ltwa-all.ldf jabbrv-ltwa-en.ldf TN Images without defects and correctly classified FP CNN classifies images as containing defects but that does not contain any defect FN CNN classifies images as not containing any defect but are containing defect Several measures are used to evaluate the efficacy of the suggested model. These include the confusion matrix, precision, recall, f1-score, and the Receiver Operating Characteristic (ROC) curve. You can figure out your F1-score, Accuracy, Precision, and Recall by solving the corresponding equations (1), (2), (3), and (4). Accuracy=\(\frac{TP+TN}{TP+FN+TN+FP}\) (1) Precision=\(\frac{\text{TP}}{TP+FP}\) (2) Recall=\(\frac{\text{TP}}{TP+FN}\) (3) F1-Score=\(2\times\frac{Precision\times Recall}{Precision+Recall}\)(4) The ROC curve shows how the true positive rate (Sensitivity) relates to the false positive rate (Specificity) at varying cutoff levels. Unlike the macro-average ROC curve, which calculates values individually for each class and then takes an average, the micro-average ROC curve combines individual values for numerous classes to compute the average. One way to measure a model’s discriminatory power is by calculating its area under the receiver operating characteristic (AUC) curve. An AUC of 1 indicates faultless performance on the evaluation of test cases. Results and Discussion The proposed deep CNN model was trained using both a training set and an evaluation set. Figure5 displays the accuracy and loss curves over training epochs as well as test results. The experimental study demonstrates that with epochs of 10, the suggested model can get a very high accuracy of 96.26% in the training set and 93.27% in the test set. With training losses of 0.096 and testing losses of 0.172, the suggested model is able to accurately categorize steel surface flaws. For the given hyperparameters, the graphs indicate that as the number of epochs increases, the accuracy of both the training and validation sets progressively increases over time, until leveling out. jabbrv-ltwa-all.ldf jabbrv-ltwa-en.ldf Figure 5. Graphical representation of accuracy and loss curve of the proposed model. Here we give the tabulated outcomes of a test conducted on a dataset designed to detect defects on steel surfaces.The confusion matrix for the proposed CNN model is shown in Figure6; a total of 4134 pictures were utilized for evaluation. When compared to other classes, the model’s identification of class 3scaledefect is particularly accurate, at 99%. The model also shows promising results when used to the prediction of additional groups of steel surface defects. A bespoke dataset of four types of images (rollmarks, scales, scratches, and normal) has been curated for the proposed research topic. Both intra-class (flaws of the same kind) and inter-class (defects of different types) defects are included in this collection. Figure 1 shows instances of intra-class pictures, which are more common in the dataset, with the scale flaws displayed in the third column. Rollmark, scratch, scale, and normal photos all make up the inter-class samples. In conclusion, the sixth difficult issue indicated in Section 1 is addressed by the suggested surface defect database. The testing accuracy of our suggested model was 93.27%, proving its capacity to deal with such formidable difficulties. Figure 6.Representation of Confusion Matrix Table 4. displays various performance measures used to evaluate the model. Precision, recall, f1-score, and support are utilized to assess its performance. The model achieves precision rates of 96% for Class normal, 90% for Class rollmark, and 99% for Class scale and 88% for scratches, resulting in an average precision of 93.33%. Table 4.Classification Report Normal 0.96 0.91 0.93 877 Rollmark 0.90 0.99 0.94 697 Scale 0.99 0.88 0.94 1427 Scratches 0.88 0.97 0.92 1133 jabbrv-ltwa-all.ldf jabbrv-ltwa-en.ldf CNN has been used as an end-to-end learning framework by several authors [16, 19]. When compared to the Deep CNN model presented by Jonathan Masci et al. [16], ours performs marginally better. The accuracy of the proposed model is compared to that of existing Deep CNN models in Table 5. Table 5. A comparative analysis of the proposed model with other Deep CNN models with respect to accuracy Max-Pooling CNN by Jonathan Masci et al.[16] Deep CNN classification 93.03% PLCNN by Gao Yiping et al.[19] Deep CNN classification 90.07% Proposed Model Deep CNN classification 93.27% Figure 7 depicts the ROC curve, a graphical representation of the model’s performance. ROC curves, or receiver operating characteristic curves, are graphs that demonstrate the trade-off between the true positive rate (TPR) and the false positive rate (FPR) for varying threshold values between 0 and 1. The proportion of positive samples that were accurately recognized as positive is known as the True Positive Rate (TPR), whereas the proportion of negative samples that were incorrectly classified as positive is known as the False Positive Rate (FPR). Receiver operating characteristic (ROC) curves provide a graphical representation of the performance of classifiers at varying thresholds. Here, the effectiveness of the classifier is determined by the area under the ROC curve (AUC). The AUC is a fraction of the area of a square unit, and it ranges from 0 to 1. A higher AUC value indicates better classification ability, as proposed by Fawcett et al. [24]. In addition to focusing on the test with the best classification performance (100 percent), the study calculates the area under the curve (AUC) and receiver operating characteristic (ROC) curves for each class. The study gives a thorough evaluation of the classifier’s classification performance and allows for comparisons across multiple thresholds and classes by using ROC curves and AUC values. Using the strength of a deep CNN model, we provide a complete approach to surface defect picture classification and diagnosis. To eliminate the requirement for a separate feature extraction phase, our novel method employs deep Convolutional Neural Networks (CNNs) to extract defect features and categorize defect categories concurrently. Our carefully curated surface defect dataset yields a remarkable 93.27% defect detection accuracy using the suggested technique. We use a number of different data augmentation strategies to make the model more robust and less prone to overfitting. In the future, we plan to increase the number of fault types we have for steel strip surfaces, providing a more complex and rigorous testing ground. We also intend to investigate more complex network designs to boost the CNN model’s fault identification capabilities. We anticipate that these next efforts will enhance the practicality and efficacy of our approach. Figure7.ROC (Receiver Operating Characteristic) curve of the proposed deep CNN model. Conclusion Flaw detection is essential for ensuring the quality, safety, and efficiency of steel production. There are a number of different approaches to Automatic Surface Inspection (ASI), but they often need some level of expertise on the part of the user. It’s a difficult problem to solve to increase ASI accuracy without using any prior information. To this end, we present a brand new nine-layer convolutional model for identifying flaws in digital photographs of steel strips. This model processes a wide range of inputs in real time, such as vibrations, light, and blur. It can reach an accuracy of up to 93.27% in fault identification without the use of a separate feature extraction phase. We investigate strategies for augmenting data to avoid overfitting. A thorough training and testing procedure is used to measure the model’s performance. Its success depends on constant tuning based on new facts from the actual world. A potential alternative for defect-free goods and enhanced production processes, this deep CNN-based technique considerably raises manufacturing quality requirements.Introducing our end-to-end surface defect identification technology might ensure fault-free goods and improve manufacturing operations. We improve industry standards and production outcomes by effortlessly integrating deep learning into ASI. Acknowledgment: The authors express their gratitude to JSW Steel Ltd,Vijayanagara, Ballari, Karnataka, India, for generously providing the dataset used in this study. References 1. NirbharNeogi, Dusmanta K Mohanta, Pranab K Dutta. Review of vision-based steel surface inspection systems. EURASIPJ.ImageVideo Process, 50,2014. 2. Nong Jin, Shiyu Zhou, Tzyy-Shuh Chang. Identification of impacting factors of surface defects in hot rolling processes using multi-level regression analysis. Southfield, MI, USA: Society of Manufacturing Engineers. (2000). 3. P.Muñoz-Escalona,A.Shokrani,S.T.Newman.Influence of cutting environments on surface integrity and power consumption of austenitic stainless steel. Robotics and Computer-Integrated Manufacturing , 36 ,pp: 60-69, (2015). 4. Jinrui Gan, Jianzhu Wang, Haomin Yu, Qingyong Li, Zhiping Shi. Online rail surface inspection utilizing spatial consistency and continuity. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 50 (7),pp:L2741-2751,(2018). 5. Matti Niskanen, Olli Silvén, HannuKauppinen. Color and texture based wood inspection with non-supervised clustering. Proceedings of the scandinavian Conference on image analysis . pp: 336-342, (2001). 6. B. Suvdaa, J. Ahn and J. Ko, Steel Surface Defects Detection and Classification Using SIFT and Voting Strategy, International Journal of Software Engineering and Its Applications , 6 (2),pp:161-166,(2012). 7. Santanu Ghorai, AnribanMukherjee,Gangadharan,Pranab.Automatic Defect Detection on Hot-Rolled Flat Steel Products. IEEETransactionsonInstrumentationand Measurement, 62 (3),pp:612-621,(2013). 8. Florent Dupont, ChristpoheOdet, Michel Carton. Optimization of the recognition of defects in flat steel products with the cost matrices theory. NDT& E International 30 (1),pp: 3–10, (1997). 9. Bae-KeunKwon,Jong-Seob Won, Dong-Joong Kang. Fast defect detection for various types of surfaces using random forest with VOV features. International Journal of Precision Engineering and Manufacturing , 16 (5),pp:965-970, (2015). 10. K.V.Prasad, Hanumesh Vaidya, Kumar Swamy K, Renuka S. Pumpkin Seeds Classification: Artificial Neural Network and Machine Learning Methods. International Journal of Academy of Physical Sciences , 27 (1),pp:23-33,(2023). 11. Kerehalli Vinayaka Prasad, Hanumesh Vaidya, ChoudhariRajashekhar, Kumar Swamy Karekal, Renuka Sali. Automated neural network forecast of PM2:5 concentration, International Journal of Mathematics and Computer in Engineering, 1 (1) pp:1-12, (2023). 12. Hanumesh Vaidya,K.V.Prasad,RenukaS,KKumarSwamy.Multiclass Classification of Dry Beans using Artificial Neural Network, International Journal of Academy of Physical Sciences. 27 (2),109-124,(2023). 13. Alex Krizhevsky, Ilya Sutskever,GeofferyE. Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems , 25, (2012). 14. Karen Simonyan,Andrew Zisserman .Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv, pp:1409-1556,(2014). 15. SergeyZagoruyko, Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146 (2016). 16. Jonathan Masci, Ueli Meier, Dan Ciresan,GabrielFricout. Steel defect classification with Max-Pooling Convolutional NeuralNetworks. 2012 International Joint Conference on Neural Networks (IJCNN), pp:1-6,(2012). 17. Daniel Weimer,Bernd Scholz-Reiter, MosheShpitalni. Design of deep convolutional neural network architectures for automated feature extraction in industrial inspection. CIRP annals , 65 (1), pp: 417-420, (2016). 18. Long Wen, Xinyu Li,Liang Gao, Yuyan Zhang. A new convolutional neural network-based data-driven fault diagnosis method. IEEE Transactions on Industrial Electronics , 65 (7), pp:5990-5998, (2017). 19. Gao Yiping, Gao Liang,Li Xinyu, Yan Xuguo. A semi-supervised convolutional neural network-based method for steel surface defect recognition. Robotics and Computer Integrated Manufacturing , 61, pp:101825,(2020). 20. Yan LeCun, Yoshua Bengio. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361 (10),(1995). 21. Y-Lan Boureau,JeanPonce, YannLeCun. A theoretical analysis of feature pooling in visual recognition. Proceedings of the 27th international conference on machine learning (ICML-10) pp:111-118,(2020). 22. Xavier Bouthillier, Kishore Konda, Pascal Vincent, Roland Memisevic.Dropout as data augmentation,ArXivPrepr. ArXiv1506.08700. (2015). 23. Fawcett, T. An introduction to ROC analysis. Pattern recognition letters , 27 (8), pp:861-874,(2006). Warnakulasuriya, S., Newell W. Johnson, and I. Van der Waal (2007): Nomenclature and classification of potentially malignant disorders of the oral mucosa, Journal of oral pathology & medicine Vol.36 (10), pp: 575-580. Information & Authors Information Version history V1 Version 1 06 February 2025 Copyright This work is licensed under a Non Exclusive No Reuse License. Keywords deepcnn intra class and inter-class rollmark scale scratches and normal steel surface defect Authors Affiliations Y. Shobha 0009-0007-5527-194X [email protected] Vijayanagara Sri Krishnadevaraya University Bellary View all articles by this author Metrics & Citations Metrics Article Usage 375 views 129 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Y. Shobha. Classification and Diagnosis of Defects in Steel Surfaces Using Deep Convolutional Neural Networks. Authorea . 06 February 2025. DOI: https://doi.org/10.22541/au.173882390.08755910/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.173882390.08755910/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ff5215cdc33e2c5',t:'MTc3OTM4Mjc1Mg=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();