Steel strip surface defect detection method and device
By improving the Yolo algorithm and combining DenseNet and SE attention modules, the problem of low accuracy in detection of steel strip surface defects is solved, and high-precision detection and classification of steel strip surface defects is achieved, which improves detection efficiency and accuracy.
Patent Information
- Application Number
- CN202510134870.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-07
- Publication Date
- 2025-06-06
AI Technical Summary
The prior art is difficult to effectively detect and identify steel belt surface defects, such as pits, curls and flash seams, resulting in a decrease in the surface quality of the steel belt, affecting aesthetics and service performance.
The M2S improved Yolo algorithm is used to combine DenseNet and SE attention modules to achieve accurate constituent and classification of areas of interest for steel strip surface defects. This method improves detection accuracy through the cross-scale aggregation module and the dual relationship module, and enhances the ability to distinguish nuances through the SE attention module.
It significantly improves the accuracy of steel belt wrap defect detection, can seamlessly connect detection and classification, realize automated detection, reduce human intervention, and improve detection efficiency and accuracy.
Smart Images
Figure CN120107170A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer vision and deep learning technology, and specifically provides a method and device for detecting surface defects of a steel strip. Background Art
[0002] The following defects are prone to occur in the steel strip wrapping process due to process problems or external forces: pits, curling and flash seams. Pits are the unevenness caused by external forces on the steel strip wrapping or improper production processes during the production process. They will directly affect the surface quality of the steel strip, reduce its appearance and performance, and may also cause defects such as holes, cracks, delamination in subsequent processing, and even damage the lining layer.
[0003] Curling is the phenomenon that the edge of the steel strip curls or warps, which will directly affect the appearance quality of the product and reduce the market competitiveness of the product. It may also cause the steel strip to break, delaminate, or even damage the lining layer during subsequent processing or use.
[0004] Flash seams refer to gaps or voids that appear between steel belts or between steel belts and cables during the steel belt wrapping process. The causes may be pitch gear mismatch, too narrow steel belt width, actual cable outer diameter exceeding the standard, and asynchronous take-up and traction causing the steel belt gap of the cable to start. This will directly damage the flatness and smoothness of the cable and reduce its appearance quality. It may also cause the skin effect and proximity effect of the cable to weaken, thereby reducing its electrical performance. It is also susceptible to erosion and damage from the external environment, thereby increasing the risk of cable failure and shortening its service life. Summary of the invention
[0005] The present invention aims at solving the above-mentioned deficiencies of the prior art and provides a method for detecting surface defects of steel strips with strong practicability.
[0006] A further technical task of the present invention is to provide a steel strip surface defect detection device that is rationally designed, safe and applicable.
[0007] The technical solution adopted by the present invention to solve the technical problem is:
[0008] A method for detecting surface defects of steel strips is proposed. Firstly, the M2S improved Yolo algorithm is used to delineate the defect region of interest. Then, the SE attention module is added on the basis of the DenseNet algorithm to enhance the ability to distinguish subtle differences and improve the accuracy of steel strip wrapping defect detection.
[0009] Furthermore, in realizing the delineation of defect areas of interest, it includes:
[0010] (1) Automatic data collection;
[0011] (2) Multiple data enhancements;
[0012] (3) The enhanced image is input into the Yolo model improved by M2S to realize the delineation of the region of interest.
[0013] Furthermore, in step (1), a data set is collected, the resolution of the image is determined, the data set is divided into a training set and a validation set, and the real-time collected data is used for testing.
[0014] Furthermore, in step (2), after automatically collecting data, any combination of random cropping, color dithering distortion and horizontal flipping is performed in the data channel to improve data diversity.
[0015] Furthermore, in step (3), the Yolo model adds a cross-scale aggregation module CAM after Backbone. The CAM consists of multiple cross-scale fusion nodes, allowing the top and bottom information to interact;
[0016] The cross-scale aggregation module CAM is connected to the head network through the dual relationship module DRM. The dual relationship module DRM captures multi-dimensional relations from CAM. Each of High, Mid and Low is fed into the module relative to DRM to obtain richer context information;
[0017] DRM includes a channel relationship module CRM and a spatial relationship module SRM. CRM processes high-dimensional features, SRM aggregates mid- and low-latitude features, and finally aggregates the high-, mid-, and low-latitude features processed by DRM to improve detection accuracy.
[0018] Furthermore, the SE attention module is added, including:
[0019] A. Design a proportional resize;
[0020] B. Use any combination of color dithering, horizontal flipping, and cutout in the data channel;
[0021] C. DenseNet+SE is an improved network based on DenseNet, in which a dense block is composed of a series of densely connected layers. As the number of layers increases, the number of feature maps generated gradually decreases, reducing the amount of calculation. The SE attention module is added to each dense block to make the extracted feature map more focused;
[0022] D. The optimizer is stochastic gradient descent SGD.
[0023] Furthermore, in step (A), if the batch size of the training process is greater than 1, the resolution of the input images needs to be the same. A proportional resizing is designed. First, the edge with the largest resize ratio is found, and the deformation ratio is calculated based on the edge with the largest ratio. The segmented image is resized, and then the edges that are less than the maximum ratio are filled with black edges.
[0024] Furthermore, in step (B), the Cutout operation is to randomly cut out the region in the image, wherein the number of the cut regions is randomly 1-16, and the length and width of the cut regions are randomly 1-32. Cutout destroys the integrity of the image to achieve a regularization effect;
[0025] In step (C), the SE attention module performs global average pooling on each feature map in the feature layer to obtain a vector with a length of the number of channels c. Then, it is input into the fully connected network to calculate the weight of each feature map. Finally, it is multiplied with the input feature layer to obtain the feature layer after weight optimization.
[0026] Furthermore, in step (D), the output is a vector of length 4, which are background, pit, curling edge and flash seam, and the position of the maximum value is taken as the output category.
[0027] A steel strip surface defect detection device, comprising: at least one memory and at least one processor;
[0028] The at least one memory is used to store a machine-readable program;
[0029] The at least one processor is used to call the machine-readable program to execute a method for detecting surface defects of a steel strip.
[0030] Compared with the prior art, the method and device for detecting surface defects of a steel strip of the present invention have the following outstanding beneficial effects:
[0031] In the defect area of interest delineation stage, the present invention uses Yolo to delineate the area of interest of steel strip wrapping, introduces M2S (Multi-to-Single) into the Yolo network, and uses a cross-scale aggregation module (CAM) and a dual relationship module (DRM) to improve the accuracy of small target defect delineation.
[0032] The SE (Squeeze-and-Excitation) attention module is introduced in the classification stage to significantly improve the network model's attention to useful features, thereby effectively solving the problem of small gaps between defect classes.
[0033] The two stages of detection and classification are seamlessly connected, so that in the reasoning stage, only the original image needs to be input to directly output the circled defect category, without the need for human intervention throughout the process. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0035] Attached Figure 1 It is a structural schematic diagram of an improved Yolo model in a steel strip surface defect detection method;
[0036] Attached Figure 2 It is a schematic diagram of the structure of a cross-scale fusion node in a steel strip surface defect detection method;
[0037] Attached Figure 3 It is a structural schematic diagram of a dual relationship module DRM in a steel strip surface defect detection method;
[0038] Attached Figure 4 It is a structural diagram of a dense block in a method for detecting surface defects of a steel strip;
[0039] Attached Figure 5 It is a structural schematic diagram of the SE attention module in a steel strip surface defect detection method. DETAILED DESCRIPTION
[0040] In order to enable those skilled in the art to better understand the solution of the present invention, the present invention is further described in detail below in conjunction with specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0041] A best embodiment is given below:
[0042] In this embodiment, a method for detecting surface defects of steel strips is provided. First, the M2S improved Yolo algorithm is used to delineate the defect region of interest. Then, the SE attention module is added to the DenseNet algorithm to enhance the ability to distinguish subtle differences and improve the accuracy of steel strip wrapping defect detection.
[0043] In the implementation of defect area of interest delineation, it includes:
[0044] (1) Automatic data collection;
[0045] A total of 2500 data sets were collected with a resolution of 1280×1280, including 2000 training sets and 500 validation sets. The data were collected in real time for testing.
[0046] (2) Multiple data enhancements;
[0047] The 2D camera automatically collects data. The images collected by cables with different wire diameters are different. To eliminate the above influence, any combination of data enhancement methods such as random cropping, color dithering distortion, and horizontal flipping are performed in the data channel (except for color dithering distortion, all other methods are processed simultaneously with the original image and label) to improve the diversity of the data.
[0048] (3) The enhanced image is input into the Yolo model improved by M2S to delineate the region of interest;
[0049] like Figure 1 As shown in Figure 2, the improved Yolo model adds a cross-scale aggregation module (CAM) after Backbone, as shown in Figure 2. Figure 2 As shown in the figure, CAM is composed of multiple cross-scale fusion nodes (CFN), which better integrates the bottom-up features of the backbone network while allowing the top and bottom information to interact. The semantics of features at different scales are integrated to improve the feature extraction ability of the model.
[0050] The cross-scale aggregation module CAM is connected to the head network through the dual relation module DRM, which captures multi-dimensional relations from CAM to enhance and calibrate the input. Each of High, Mid, and Low is fed into the module relative to DRM to obtain richer contextual information.
[0051] like Figure 3 ,DRM includes a channel relationship module (CRM) and a spatial relationship module (SRM), in which CRM processes high-dimensional features, SRM aggregates mid- and low-dimensional features, and finally aggregates the high-, mid-, and low-dimensional features processed by DRM, achieving better detection accuracy improvement.
[0052] In the SE attention module, including:
[0053] A. Design a proportional resize;
[0054] A total of 2500 data sets were collected, and the original image resolution was 1280×1280. However, after the Stagone defect region of interest was determined, the image resolution was uncertain. If the batch in the training process was greater than 1, the input image resolution needed to be the same (the resolution used in the present invention was 224×224). If the image was directly resized, it would cause deformation, which was not conducive to the classification of defects in the region of interest of steel strip wrapping. A proportional resize was designed. First, the edge with the largest ratio to 224 was found, and the deformation ratio was calculated based on the edge. The segmented image was resized, and the edges less than 224 were filled with black edges.
[0055] B. Use any combination of color dithering, horizontal flipping, and cutout in the data channel;
[0056] In the data channel, any combination of data enhancement methods such as color jitter, horizontal flip, and Cutout are used. The Cutout operation randomly cuts off the area in the image, where the number of cut areas is 1-16 randomly, and the length and width of the cut area are both 1-32 randomly.
[0057] Since the entire cable area is input into the classification network, but the cable area that can truly determine the cable status classification accounts for a very small part of the total area, Cutout can destroy the integrity of the image and achieve a regularization effect.
[0058] C. DenseNet+SE is an improved network based on DenseNet, such as Figure 4 As shown in the figure, the denseblock is composed of a series of densely connected layers. As the number of layers increases, the number of feature maps generated gradually decreases, reducing the amount of calculation;
[0059] like Figure 5 As shown in the figure, the SE attention module takes the global average pooling of each feature map in the feature layer to obtain a vector with a length of the number of channels c, and then inputs it into the fully connected network to calculate the weight of each feature map, and finally multiplies it with the input feature layer to obtain the weight-optimized feature layer. SE dense block optimizes the weight of the output feature layer of the dense block.
[0060] D. Loss is bceloss, and the optimizer is stochastic gradient descent SGD;
[0061] The output is a vector of length 4, which are background, pits, curling edges and flash seams. The position of the maximum value is taken as the output category.
[0062] Based on the above method, a steel strip surface defect detection device in this embodiment includes: at least one memory and at least one processor;
[0063] The at least one memory is used to store a machine-readable program;
[0064] The at least one processor is used to call the machine-readable program to execute a method for detecting surface defects of a steel strip.
[0065] The above-mentioned specific implementations are only specific cases of the present invention. The patent protection scope of the present invention includes but is not limited to the above-mentioned specific implementations. Any technical solutions that conform to the above-mentioned specific implementations of the present invention and any appropriate changes or substitutions made by ordinary technicians in the relevant technical field shall fall within the patent protection scope of the present invention.
[0066] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for detecting surface defects of a steel strip, characterized in that: Firstly, the M2S improved Yolo algorithm is used to delineate the defect area of interest; Then, the SE attention module is added to the DenseNet algorithm to enhance the ability to distinguish subtle differences and improve the accuracy of steel strip wrapping defect detection.
2. A method for detecting surface defects of a steel strip according to claim 1, characterized in that: In the implementation of defect area of interest delineation, it includes: (1) Automatic data collection; (2) Multiple data enhancements; (3) The enhanced image is input into the Yolo model improved by M2S to realize the delineation of the region of interest.
3. A method for detecting surface defects of a steel strip according to claim 2, characterized in that: In step (1), a data set is collected, the resolution of the image is determined, the data set is divided into a training set and a validation set, and the real-time collected data is used for testing.
4. A method for detecting surface defects of a steel strip according to claim 3, characterized in that: In step (2), after automatically collecting data, any combination of random cropping, color dithering distortion, and horizontal flipping is performed in the data channel to improve data diversity.
5. A method for detecting surface defects of a steel strip according to claim 4, characterized in that: In step (3), the Yolo model adds a cross-scale aggregation module CAM after Backbone. CAM consists of multiple cross-scale fusion nodes, allowing top and bottom information to interact; The cross-scale aggregation module CAM is connected to the head network through the dual relationship module DRM. The dual relationship module DRM captures multi-dimensional relations from CAM. Each of High, Mid and Low is fed into the module relative to DRM to obtain richer context information; DRM includes a channel relationship module CRM and a spatial relationship module SRM. CRM processes high-dimensional features, SRM aggregates mid- and low-latitude features, and finally aggregates the high-, mid-, and low-latitude features processed by DRM to improve detection accuracy.
6. A method for detecting surface defects of a steel strip according to claim 5, characterized in that: In the SE attention module, including: A. Design a proportional resize; B. Use any combination of color dithering, horizontal flipping, and cutout in the data channel; C. DenseNet+SE is an improved network based on DenseNet, in which a dense block is composed of a series of densely connected layers. As the number of layers increases, the number of feature maps generated gradually decreases, reducing the amount of calculation. The SE attention module is added to each dense block to make the extracted feature map more focused; D. The optimizer is stochastic gradient descent SGD.
7. A method for detecting surface defects of a steel strip according to claim 6, characterized in that: In step (A), if the batch size of the training process is greater than 1, the resolution of the input images needs to be the same. A proportional resizing is designed. First, the edge with the largest resize ratio is found. The deformation ratio is calculated based on the edge with the largest ratio. The segmented image is resized, and the edges that are less than the maximum ratio are filled with black edges.
8. A method for detecting surface defects of a steel strip according to claim 7, characterized in that: In step (B), the Cutout operation is to randomly cut out the area in the image, wherein the number of the cut areas is randomly set between 1 and 16, and the length and width of the cut areas are randomly set between 1 and 32. Cutout destroys the integrity of the image to achieve a regularization effect; In step (C), the SE attention module performs global average pooling on each feature map in the feature layer to obtain a vector with a length of the number of channels c. Then, it is input into the fully connected network to calculate the weight of each feature map. Finally, it is multiplied with the input feature layer to obtain the feature layer after weight optimization.
9. A method for detecting surface defects of a steel strip according to claim 8, characterized in that: In step (D), the output is a vector of length 4, which are background, pit, curling edge and flash seam respectively, and the position of the maximum value is taken as the output category.
10. A steel strip surface defect detection device, characterized in that: include: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to call the machine-readable program to execute the method according to any one of claims 1 to 9.