Metal surface defect detection method based on improved YOLOv5

By adding the CBAM attention mechanism in the YOLOv5 model, modifying the convolutional layer to GSConv, and using the EIoU Loss border regression loss function, the problems of low detection accuracy and long time in metal surface defect detection are solved, achieving more efficient and accurate Detection effect.

CN119067919AInactive Publication Date: 2024-12-03SHENYANG UNIVERSITY OF TECHNOLOGY

Patent Information

Application Number
CN202411042802.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-12-03
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing metal surface defect detection algorithms are difficult to balance between detection accuracy and real-time, and the defect feature extraction capability is insufficient and the detection time is long.

Method used

Based on the YOLOv5 model, the CBAM attention mechanism was added, the convolution layer of the Neck layer was modified to GSConv, and the EIoU Loss border regression loss function was used to optimize the network structure to improve detection accuracy and efficiency.

Benefits of technology

It significantly improves detection accuracy and efficiency, shortens detection time, and improves the overall accuracy and real-timeness of metal surface defect detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067919A_ABST
    Figure CN119067919A_ABST
Patent Text Reader

Abstract

The invention discloses a metal surface defect detection method based on improved YOLOv5, and provides a metal surface defect detection model based on an improved YOLOv5 algorithm aiming at the problems of insufficient defect feature extraction capability, low detection precision, long detection time and the like in existing metal surface defect detection. Firstly, a CBAM attention mechanism is added before an SPPF module, a network structure is optimized, and richer target features are extracted; secondly, the common convolution Conv in the Neck layer is modified into GSConv for training, model parameters are reduced, and the training speed is increased; and finally, introducing a frame regression loss function EIoU Loss to replace CIoU Loss of the original YOLOv5 to accelerate network convergence and increase the small target detection capability. According to the invention, the detection precision is obviously improved, the time required for detection is successfully shortened, and the overall efficiency and accuracy of metal surface defect detection are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field related to image processing, and in particular to a metal surface defect detection method based on improved YOLOv5. Background Art

[0002] As an important raw material for industry, metal is widely used in the production fields of machinery, aerospace, automobile, national defense, light industry, etc. However, due to the influence of factors such as raw material quality, production environment, equipment, and human errors, various problems may occur on the surface of the produced metal. The generation of metal surface defects can easily damage the surface of the steel, reduce its quality, and thus affect the use effect and life of the metal. Therefore, metal surface defect detection is widely regarded as a crucial link in the production process.

[0003] At present, deep algorithms represented by convolutional neural networks have made significant progress in the field of computer vision, and have been widely used in the field of target detection of metal surface defects. At present, target detection algorithms are divided into two categories: one is a one-stage target detection algorithm represented by the YOLO series, SSD, etc., which combines candidate area division and object detection. This algorithm has a fast detection speed, but there is still a lot of room for improvement in detection accuracy, and the detection ability for small targets is weak; the other is a two-stage target detection algorithm represented by Mask R-CNN, Faster R-CNN, etc. Although the recognition accuracy of the two-stage target detection algorithm is relatively high, its detection process is divided into two steps, the detection speed is slow, and it is not possible to perform real-time detection well. Therefore, the mainstream target algorithm at this stage cannot balance the accuracy and real-time performance of detection, and there are still great limitations in the application of metal surface defect detection. Summary of the invention

[0004] The present invention aims to solve the problems of insufficient defect feature extraction capability, low detection accuracy, long detection time, etc. of the current metal surface defect detection algorithm. Improvements are made on the basis of the original YOLOv5 model, and a metal surface defect detection method based on improved YOLOv5 is proposed. The improved model not only improves the detection accuracy, but also effectively reduces the number of model parameters, thereby improving the defect detection efficiency.

[0005] In order to achieve the above object, the present invention provides a metal surface defect detection method based on improved YOLOv5, comprising the following steps:

[0006] The steps include:

[0007] S10. Collecting metal surface defect images and constructing a data set of metal surface defects;

[0008] S20. Preprocessing the data set of the metal surface and dividing the data set;

[0009] S30. Build a metal surface defect model based on the improved YOLOv5, add the CBAM attention mechanism before the SPPF in the Backbone layer of the YOLOv5 network, modify the ordinary convolution Conv in the Neck layer to GSConv, and use the bounding box regression loss function EIoU Loss instead of the CIoU Loss of the original YOLOv5;

[0010] S40. Train the improved YOLOv5 model to achieve recognition;

[0011] S50. Evaluate the YOLOv5 metal surface defect detection model after training.

[0012] In some modified embodiments of the first aspect of the present invention, the defect category statistics of the data set include cracks, plaques, inclusions, pressed oxide scale, scratches and pitting surfaces, and the sample images of each defect category

[0013] In some modified implementations of the first aspect of the present invention, in step S20, the data set of the metal surface is divided into a training set and a test set according to a ratio of 9:1, and the images of the metal surface defect data set and their corresponding label files are divided.

[0014] In some modified implementations of the first aspect of the present invention, in step S30, the CBAM attention mechanism includes a channel attention module and a spatial attention module.

[0015] The channel attention mechanism performs global maximum pooling and global average pooling on the input feature map to describe the characteristics of each channel, and then passes the result to the multi-layer perceptron for processing. Finally, the channel attention weight is formed by the Sigmiod function and multiplied with the input feature map F to obtain the weighted feature map M. C , the expression is:

[0016]

[0017] Among them, M C (F) represents the output of the channel attention mechanism module; δ represents the Sigmiod function; MLP represents multi-layer perceptron; AvgPool represents average pooling; MaxPool represents maximum pooling; W1 and W0 represent MLP weight matrices; and Denotes the average pooled feature vector and the maximum pooled feature vector of the channel respectively;

[0018] The spatial attention mechanism outputs the feature map M with the channel attention mechanism. CThe input features are global maximum pooling and global average pooling between channels. After channel splicing, 7×7 convolution is performed. Finally, the spatial attention weight is generated by the Sigmoid function and multiplied with the input feature map to obtain the weighted feature map M. S , whose expression is:

[0019]

[0020] Among them, M S (F) represents the output of the spatial attention mechanism module; δ represents the sigmiod function; F 7×7 Represents a 7×7 convolution operation; AvgPool represents average pooling; MaxPool represents maximum pooling; and Represents the average pooled feature vector and the maximum pooled feature vector of the space respectively.

[0021] In some modified implementations of the first aspect of the present invention, in S30, when the ordinary convolution Conv in the Neck layer of the original YOLOv5 network is modified to GSConv, the outputs of the ordinary convolution layer SC and the depthwise separable convolution layer DSC are spliced, and the spliced ​​feature maps are shuffled.

[0022] In some modified implementations of the first aspect of the present invention, in step 30, the bounding box regression loss function EIoU Loss is used to replace the CIoU Loss of the original YOLOv5, and its expression is:

[0023]

[0024] Among them, IoU is the intersection over union ratio of the real box and the predicted box; ρ(b,b gt ) is the Euclidean distance between the two center points; b and b gt is the center point of the predicted box and the real box; c is the diagonal distance of the minimum circumscribed matrix between the predicted box and the real box; w and h represent the width and height of the predicted box respectively; w gt and h gt are the width and height of the real box respectively. w and C h are the width and height of the minimum circumscribed matrix of the predicted box and the true box, respectively.

[0025] In some modified implementations of the first aspect of the present invention, in step S40, five metal surface defect model ablation experiments obtained during model training are compared; they are original YOLOv5 model training experiment, YOLOv5+CBAM model experiment, YOLOv5+GSConv model experiment, YOLOv5+EIoU and improved YOLOv5 model experiment, each model is trained for 200 rounds, and the image size is 640×640.

[0026] In some modified implementations of the first aspect of the present invention, in step S50, the evaluation indicators for the improved YOLOv5 model mainly include precision, recall, mean average precision mAP, and detection rate FPS; the precision P is used to evaluate the accuracy of the metal surface defect detection; the recall R is used to evaluate the comprehensiveness of the model metal surface defect detection; mAP is a comprehensive indicator used to evaluate the performance of the target detection model, which is the average of the average precision of all categories; FPS represents frames per second; the expression is as follows:

[0027]

[0028] Among them, TP means that the metal surface defect detection model successfully identifies the true positive samples; FP means that the metal surface defect detection model misclassifies the actual negative samples as positive; FN means that the metal surface defect detection model mistakenly classifies the positive samples as negative;

[0029]

[0030] Wherein, AP represents the area of ​​the RP curve; n represents the number of detection categories.

[0031] The beneficial effects of the present invention are as follows:

[0032] In view of the problems of insufficient defect feature extraction capability, low detection accuracy, long detection time, etc. in the existing metal surface defect detection, the present invention proposes a metal surface defect detection model based on the improved YOLOv5 algorithm. First, the CBAM attention mechanism is added before the SPPF module to optimize the network structure and extract richer target features. Secondly, the ordinary convolution Conv in the Neck layer is modified to GSConv for training to reduce model parameters and improve the training speed. Finally, the bounding box regression loss function EIoU Loss is introduced to replace the CIoU Loss of the original YOLOv5 to speed up network convergence and increase the small target detection capability. The present invention achieves a significant improvement in detection accuracy, successfully shortens the time required for detection, and improves the overall efficiency and accuracy of metal surface defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The improved YOLOv5 structure diagram of the present invention;

[0034] Figure 2 Schematic diagram of the CBAM attention mechanism of the present invention;

[0035] Figure 3 The channel attention principle diagram of the present invention;

[0036] Figure 4 Schematic diagram of the spatial attention principle of the present invention;

[0037] Figure 5 Schematic diagram of GSConv of the present invention;

[0038] Figure 6 The EIoU Loss principle diagram of the present invention;

[0039] Figure 7 Comparison of YOLOv5 visualization effects before and after the improvement of the present invention. DETAILED DESCRIPTION

[0040] In order to better illustrate the purpose, technical solutions and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be noted that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0041] Embodiment 1

[0042] like Figure 1 The present invention proposes a metal surface defect detection method based on improved YOLOv5, comprising the following steps:

[0043] Step S10. Collect metal surface defect images and construct a metal surface defect data set;

[0044] The present invention uses the NEU-DET metal surface defect dataset, which includes six typical metal surface defects, namely cracks, plaques, inclusions, pressed oxide scale, scratches, and pitting surfaces;

[0045] The dataset contains 1800 images, and each defect category preferably has 300 samples.

[0046] Step S20. Preprocessing the metal surface data set and dividing the data set;

[0047] The metal surface dataset is divided into training set and test set according to the ratio of 9:1, so the training set has 1620 images and the test set has 180 images;

[0048] Divide the metal surface defect dataset images and their corresponding label files Labels.

[0049] Step S30. Build a metal surface defect model based on the improved YOLOv5, add a CBAM attention mechanism before the SPPF in the Backbone layer of the YOLOv5 network, modify the ordinary convolution Conv in the Neck layer to GSConv, and use the border regression loss function EIoU Loss instead of the CIoU Loss of the original YOLOv5;

[0050] Add the CBAM attention mechanism before the SPPF module in the Backbone layer of the YOLOv5 network, such as Figure 2 It consists of a channel attention module and a spatial attention module, which can intelligently assign weights to the features of the input image, so that the model can adaptively focus on the most critical information of each feature, thereby improving the overall performance of the model.

[0051] like Figure 3 The channel attention mechanism performs global maximum pooling and global average pooling on the input feature map to describe the characteristics of each channel, and then passes the result to the multi-layer perceptron for processing. Finally, the channel attention weight is formed by the Sigmiod function and multiplied with the input feature map F to obtain the weighted feature map M. C , the formula is:

[0052]

[0053] Among them, M C (F) represents the output of the channel attention mechanism module; δ represents the Sigmiod function; MLP represents multi-layer perceptron; AvgPool represents average pooling; MaxPool represents maximum pooling; W1 and W0 represent MLP weight matrices; and Denote the average pooled feature vector and the maximum pooled feature vector of the channel respectively.

[0054] like Figure 4 The spatial attention mechanism outputs the feature map M with the channel attention mechanism. C The input features are global maximum pooling and global average pooling between channels. After channel splicing, 7×7 convolution is performed. Finally, the spatial attention weight is generated by the Sigmoid function and multiplied with the input feature map to obtain the weighted feature map M. S , the formula is:

[0055]

[0056] Among them, M S (F) represents the output of the spatial attention mechanism module; δ represents the sigmiod function; F 7×7Represents a 7×7 convolution operation; AvgPool represents average pooling; MaxPool represents maximum pooling; and Represents the average pooled feature vector and the maximum pooled feature vector of the space respectively.

[0057] Modify the ordinary convolution Conv in the Neck layer of the original YOLOv5 network to GSConv, such as Figure 5 It mainly concatenates the outputs of ordinary convolution SC and depthwise separable convolution DSC, and performs shuffle operation on the concatenated feature maps. The shuffling process is to convolve the input feature map with C1 channels, where C1 is the number of channels of the input feature map, to generate a C2 / 2 feature vector, and then obtain another C2 / 2 feature vector through depthwise separable convolution, concatenate the two feature vectors, and finally distribute the information generated by SC to each DSC partial output feature map through channel shuffling. The number of channels is C2, achieving uniform fusion of information between channels, which can reduce the complexity of the Neck layer and effectively improve the training speed of the model.

[0058] The bounding box regression loss function EIoU Loss is used to replace the CIoU Loss of the original YOLOv5, such as Figure 6 It introduces more penalty items to accurately adjust and minimize the difference between the width and height of the target box and the anchor box, enhance the stability of the target detection network, and effectively improve the performance of the metal surface defect detection model. The formula is:

[0059]

[0060] Among them, IoU is the intersection over union ratio of the real box and the predicted box; ρ(b,b gt ) is the Euclidean distance between the two center points; b and b gt is the center point of the predicted box and the real box; c is the diagonal distance of the minimum circumscribed matrix between the predicted box and the real box; w and h represent the width and height of the predicted box respectively; w gt and h gt are the width and height of the real box respectively. w and C h are the width and height of the minimum circumscribed matrix of the predicted box and the true box, respectively.

[0061] Step S40: training the improved YOLOv5 model to achieve recognition;

[0062] The five metal surface defect model ablation experiments obtained during model training were compared; they were the original YOLOv5 model training experiment, the YOLOv5+CBAM model experiment, the YOLOv5+GSConv model experiment, the YOLOv5+EIoU model experiment, and the improved YOLOv5 model experiment. Each model was trained for 200 rounds, and the image size was 640×640. The experimental results are shown in Table 1:

[0063] Table 1: Comparison of ablation experiments of the present invention

[0064]

[0065] Analysis of ablation experiment results: Compared with the original YOLOv5 model, the improved YOLOv5 model of the present invention has improved the accuracy (Precision) by 5.1%, the recall rate (Recall) by 5.3%, the mAP by 6.2%, and the FPS by 2.1 frames in metal surface defect detection, indicating that the improved YOLOv5 model can effectively reduce model parameters, improve detection accuracy, and is conducive to model training and deployment.

[0066] Step S50: Evaluate the trained YOLOv5 metal surface defect detection model.

[0067] In step S50, the evaluation indicators for the improved YOLOv5 model mainly include precision, recall, mean average precision (mAP), and detection rate (FPS); the precision rate P is used to evaluate the accuracy of the metal surface defect detection; the recall rate R is used to evaluate the comprehensiveness of the metal surface defect detection of the model; mAP is a comprehensive indicator used to evaluate the performance of the target detection model, which is the average of the average precision of all categories; FPS stands for frames per second; the expression is as follows:

[0068]

[0069] Among them, TP means that the metal surface defect detection model successfully identifies the true positive samples; FP means that the metal surface defect detection model misclassifies the actual negative samples as positive; FN means that the metal surface defect detection model mistakenly classifies the positive samples as negative;

[0070]

[0071]

[0072] in, AP Represents the area of ​​the RP curve; n represents the number of detection categories.

[0073] The visualization comparison of the YOLOv5 model before and after improvement is shown below. Figure 7 As shown, the improved model can better distinguish the background from the target; it can accurately identify the defect category and locate more accurately; there are a large number of dense small target defects, almost all of which can be detected by the improved YOLOv5 model of the present invention, effectively solving the current problems of metal surface defects such as insufficient defect feature extraction capability, low detection accuracy, and long detection time.

[0074] It should be noted that in this embodiment, the improved YOLOv5 model is improved in three aspects: improvement of the Backbone layer, improvement of the Neck layer and improvement of the output loss function. It is mainly composed of the input end (Input), the backbone network (Backbone), the head network (Neck) and the output end (Head).

[0075] 1. Input: The input of the improved YOLOv5 model continues to use the original YOLOv5 design. Data enhancement and adaptive calculation of the aiming frame are two very important concepts in the input layer, which help improve the performance and robustness of the model. Data enhancement uses Mosaic to increase the diversity of data to help the model better generalize to different scenarios and changes to improve the stability and performance of the model. The adaptive calculation of the aiming frame uses a dynamic method to determine the size and proportion of the aiming frame to better match the size of different targets, which can improve the detection speed of the model.

[0076] 2. Backbone network: The Backbone layer of the improved YOLOv5 model mainly includes Conv, C3, and SPPF modules, which use convolutional neural networks to perform multi-level feature extraction, gradually extracting different features of the image from low to high levels. These features will be used for subsequent target detection tasks. Adding the CBAM attention mechanism before the SPPF module in the Backbone layer of the YOLOv5 network can intelligently weight the features of the input image, allowing the model to adaptively focus on the most critical information of each feature, thereby improving the overall performance of the model.

[0077] 3. Head network (Neck): The Neck layer of the improved YOLOv5 model is responsible for connecting or fusing multiple different feature maps from the backbone network to improve the performance of target detection. This helps the model capture targets and contextual information of different scales and improve detection accuracy. When the ordinary convolution Conv in the Neck layer of the original YOLOv5 network is modified to GSConv, the outputs of the ordinary convolution layer SC and the depthwise separable convolution layer DSC are spliced, and the spliced ​​feature maps are shuffled, which can reduce the complexity of the Neck layer and effectively improve the training speed of the model.

[0078] 4. Output (Head): The output of the original YOLOv5 model uses the bounding box loss function CIoU Loss as the detection head. Although CIoU Loss takes into account the overlapping area of ​​the bounding box regression function, the distance between the center positions, and the aspect ratio, CIoU Loss can only reflect the difference in aspect ratio to a certain extent. This method often cannot accurately capture the difference in actual size. Therefore, the model optimization may be hindered and the similarity cannot be effectively improved. The improved EIoU Loss method not only retains the advantages of CIoU Loss, but also introduces more penalty terms to accurately adjust and minimize the difference between the width and height of the target box and the anchor box, thereby achieving faster convergence speed and better positioning results.

[0079] The above embodiments further describe the purpose, technical solutions and beneficial effects of the present invention in detail. The above contents are only a specific example of the implementation of the present invention, and are intended to illustrate but not limit the protection scope of the present invention. Any adjustments based on the core concepts and basic principles of the present invention, the adoption of similar alternatives, and further optimization and improvement shall be included in the protection scope of the present invention.

Claims

1. A metal surface defect detection method based on improved YOLOv5, It is characterized in that The steps include: S10. Collecting metal surface defect images and constructing a data set of metal surface defects; S20. Preprocessing the data set of the metal surface and dividing the data set; S30. Build a metal surface defect model based on the improved YOLOv5, add the CBAM attention mechanism before the SPPF in the Backbone layer of the YOLOv5 network, modify the ordinary convolution Conv in the Neck layer to GSConv, and use the bounding box regression loss function EIoU Loss instead of the CIoU Loss of the original YOLOv5; S40. Train the improved YOLOv5 model to achieve recognition; S50. Evaluate the YOLOv5 metal surface defect detection model after training.

2. According to the metal surface defect detection method based on improved YOLOv5 as described in claim 1, It is characterized in that In S10, the defect category statistics of the data set include cracks, plaques, inclusions, pressed oxide scale, scratches and pitting surfaces, and the sample images of each defect category are 300 samples.

3. According to the metal surface defect detection method based on improved YOLOv5 as described in claim 1, It is characterized in that In step S20, the metal surface dataset is divided into a training set and a test set according to a ratio of 9:1, and the metal surface defect dataset images and their corresponding label files are divided.

4. According to the metal surface defect detection method based on improved YOLOv5 according to claim 1, It is characterized in that In step S30, the CBAM attention mechanism includes a channel attention module and a spatial attention module. The channel attention mechanism performs global maximum pooling and global average pooling on the input feature map to describe the characteristics of each channel, and then passes the result to the multi-layer perceptron for processing. Finally, the channel attention weight is formed by the Sigmiod function and multiplied with the input feature map F to obtain the weighted feature map M. C , the expression is: Among them, M C (F) represents the output of the channel attention mechanism module; δ represents the sigmiod function; MLP represents multi-layer perceptron; AvgPool represents average pooling; MaxPool represents maximum pooling; W 1 and W 0 represents the MLP weight matrix; and Denotes the average pooled feature vector and the maximum pooled feature vector of the channel respectively; The spatial attention mechanism outputs the feature map M with the channel attention mechanism. C The input features are global maximum pooling and global average pooling between channels. After channel splicing, 7×7 convolution is performed. Finally, the spatial attention weight is generated by the Sigmoid function and multiplied with the input feature map to obtain the weighted feature map M. S , whose expression is: Among them, M S(F) represents the output of the spatial attention mechanism module; δ represents the Sigmiod function; F 7×7 Represents a 7×7 convolution operation; AvgPool represents average pooling; MaxPool represents maximum pooling; and Represents the average pooled feature vector and the maximum pooled feature vector of the space respectively.

5. According to the metal surface defect detection method based on improved YOLOv5 according to claim 4, It is characterized in that In S30, when the ordinary convolution Conv in the Neck layer of the original YOLOv5 network is modified to GSConv, the outputs of the ordinary convolution layer SC and the depthwise separable convolution layer DSC are concatenated, and the concatenated feature maps are shuffled.

6. According to the metal surface defect detection method based on improved YOLOv5 according to claim 1, It is characterized in that In step 30, the bounding box regression loss function EIoU Loss is used to replace the CIoU Loss of the original YOLOv5, and its expression is: Among them, IoU is the intersection over union ratio of the real box and the predicted box; ρ(b,b gt ) is the Euclidean distance between the two center points; b and b gt is the center point of the predicted box and the real box; c is the diagonal distance of the minimum circumscribed matrix between the predicted box and the real box; w and h represent the width and height of the predicted box respectively; w gt and h gt are the width and height of the real frame respectively; C w and C h are the width and height of the minimum circumscribed matrix of the predicted box and the true box, respectively.

7. According to the metal surface defect detection method based on improved YOLOv5 according to claim 1, It is characterized in that In step S40, five metal surface defect model ablation experiments obtained during model training are compared; they are original YOLOv5 model training experiment, YOLOv5+CBAM model experiment, YOLOv5+GSConv model experiment, YOLOv5+EIoU and improved YOLOv5 model experiment. Each model is trained for 200 rounds, and the image size is 640×640.

8. According to the metal surface defect detection method based on improved YOLOv5 according to claim 1, It is characterized in that In step S50, the evaluation indicators of the improved YOLOv5 model mainly include precision, recall, mean average precision (mAP), and detection rate (FPS); the precision rate P is used to evaluate the accuracy of the metal surface defect detection; the recall rate R is used to evaluate the comprehensiveness of the model metal surface defect detection; mAP is a comprehensive indicator used to evaluate the performance of the target detection model. It is the average of the average precision of all categories; FPS stands for frames per second; the expression is as follows: in, TP It indicates that the metal surface defect detection model successfully identifies true positive samples; FP It means that the metal surface defect detection model misclassifies the actually negative samples as positive; FN It means that the metal surface defect model incorrectly classifies positive samples as negative; in, AP Represents the area of ​​the RP curve; n represents the number of detection categories.

Citation Information

Patent Citations

  • Lightweight seal target detection method based on YOLOv5

    CN117576373A

  • Ceramic tableware surface defect detection method based on improved YOLOv5 model

    CN117593513A

  • Intelligent detection method and unmanned surface vehicle for multiple type faults of near-water bridges

    US20230351573A1

Cited By

  • Image recognition and positioning method of component embedded part based on target detection

    CN119672717A

  • New energy lithium battery surface defect visual detection method based on YOLOv8 model

    CN120235848A

  • Metallographic defect intelligent detection system and method based on improved YOLOv11

    CN120931651A

  • Metallographic defect intelligent detection system and method based on improved YOLOv11

    CN120931651B