Multi-category defect detection method and system based on UAV distribution network overhead lines

Abstract

This invention proposes a multi-category defect detection method and system for overhead lines in an unmanned aerial vehicle (UAV) distribution network. This system belongs to the technical field of UAV inspections. By using a fine-grained sample matching method, calculating a quality alignment matrix, and designing a fine-grained compensation index to compensate for the coefficients of difficult samples in the quality alignment matrix, the quality alignment matrix is ​​normalized to address the problem of low matrix scores at the beginning of training. Furthermore, the invention proposes a dual-weight loss, using the quality alignment matrix as a weight for positive samples to improve the consistency of model classification and regression, and enhance the ability to mine difficult samples. By introducing IOU into negative samples, the classification scores of anchor frames with low IOU are reduced, thereby reducing model false positives.

Description

Technical Field

[0001] The present invention relates to the technical field of drone inspection, and in particular to a multi-category defect detection method and system based on drone-based distribution network overhead lines. Background Art

[0002] Distribution network overhead lines suffer from a wide variety of defects, with varying forms and distribution patterns. This leads to difficulties in defect detection training and low defect recall. Existing distribution network defect detection methods often employ a data pipeline combining detection and classification. However, the fusion of multiple models in a pipeline results in high false detection rates, large memory usage, and high maintenance costs.

[0003] Due to the diverse exposure environments of distribution network lines, the defects present vary in various forms, and the distribution of defect data varies greatly. For example, in the same batch of training data, there may be hundreds of thousands of samples of missing pins, while there may be only a few hundred samples of burned insulators. This makes it difficult for the model to converge, prone to gradient explosion, low recall and many false detections. The existing technology uses a pipeline form of detection and classification to fuse multiple models, which results in high false detection rates, large memory usage, and high maintenance costs. Summary of the Invention

[0004] Purpose of the invention: To propose a multi-category defect detection method and system based on drone distribution network overhead lines to solve the above-mentioned problems existing in the prior art. By processing the training sample data and training with dual-weight loss, the purpose of retaining details in the defect detection process and reducing the holes that appear in the grid reconstruction process is achieved.

[0005] Technical solution: First, a multi-category defect detection method for overhead lines in distribution networks based on drones is proposed. The method includes the following steps:

[0006] Step 1: Build a defect detection model and receive image data to be analyzed;

[0007] Step 2: extracting three image features from the image data using the defect detection model;

[0008] Step 3: Perform a multi-scale fusion operation on the three image features to obtain fusion features;

[0009] Step 4: Perform decoupling operation on the fused features to obtain regression features, classification features, and regression offset features;

[0010] Step 5: Construct a quality alignment matrix and a fine-grained compensation index, use the quality alignment matrix to obtain positive sample points, and normalize the quality alignment matrix using the fine-grained compensation coefficient.

[0011] Step 6: Construct a dual-weight loss, use the quality alignment matrix as the weight of positive samples, and eliminate ambiguous samples;

[0012] Step 7: Based on the samples after disambiguation, the final result of defect detection is obtained according to the classification features and the corrected regression features.

[0013] In some implementation methods of the first aspect, the defect detection model includes: a backbone module, a neck module, a decoupling module, a correction module and an output module. During data processing, the backbone module is first used to perform feature extraction operations to obtain three output features; secondly, the neck module is used to perform multi-scale fusion of the three output features to obtain fusion features, and the fusion features are transmitted to the decoupling module; thirdly, the decoupling module decouples the fusion results to obtain classification features, regression features and regression offset features; thirdly, the correction module is used to fuse the regression features and regression offset features to generate regression coordinates, and then obtain the specific coordinates of the detection frame; finally, the output module outputs the corrected regression features and classification features.

[0014] In some implementations of the first aspect, during the process of performing classification and regression by the defect detection model, a quality alignment matrix is ​​constructed and optimized using a fine-grained compensation coefficient, thereby performing fine-grained sample matching.

[0015] Among them, the expression of the quality alignment matrix is:

[0016] Align Metric=cls 0.5 *iou 5

[0017] Where cls represents the classification score output by the decoupling module; iou represents the intersection-over-union ratio between the regression coordinates output by the decoupling module and the true target frame;

[0018] The expression of the fine-grained compensation coefficient is:

[0019]

[0020] Where w represents the width of the true target frame; h represents the height of the true target frame; cls represents the classification score of the defect detection model; const represents the difficulty coefficient;

[0021] The expression for combining the quality alignment matrix with the fine-grained compensation coefficient is:

[0022] Align Metric′=(Align Metric) C

[0023] After obtaining the Align Metric′, a sorting operation is performed and the top 10 points are selected as positive sample points for each true target box.

[0024] After obtaining the positive sample points, the normalization operation is performed, and the corresponding expression is:

[0025] Pos_align_metrics=Align Metric′.amax(-1)

[0026] norm_Align_Metric=Align Metric′ / Pos_align_metrics

[0027] In the formula, the size of Align Metric′ is (b, obj_num, grid_num), where objnum represents the number of G real target boxes in the current image, grid_num represents the number of grids occupied by output features, and b represents the batch size; amax(-1) represents the maximum value in the dimension of grid_num; Pos_align_metrics represents the maximum value of each GT real target box in the metric quality.

[0028] In some implementations of the first aspect, the negative sample loss is decoupled from the positive sample loss by constructing a dual-weight loss. When constructing the negative sample loss, the IOU score of the anchor box is used as the evaluation coefficient, and the corresponding expression is:

[0029] neg_weight = (1-iou) 2 *cls.sigmoid()

[0030] neg_loss=BCE(cls.sigmoid(), neg_label)*neg_weight

[0031] In the formula, iou represents the intersection-over-union ratio of the regression coordinates output by the model and the true coordinate frame; cls.sigmoid() represents the sigmoid activation of the last dimension of the classification feature; BCE represents cross entropy; neg_label represents all anchor anchor frames.

[0032] The obtained normalized quality alignment matrix is ​​used as the fine-grained weight for classifying positive samples and regression loss. The corresponding expression is:

[0033] pos_weight = (1-cls.sigmoid) 2 *pos_label

[0034] pos_loss=BCE(cls.sigmoid(), norm_Align_Metric)*pos_weight

[0035] loss = neg_loss + pos_loss

[0036] Where pos_label represents the positive sample anchor box; neg_loss represents the negative sample anchor box.

[0037] Secondly, a multi-category defect detection system based on UAV distribution network overhead lines is proposed. The system includes: a model construction module, a feature extraction module, a feature fusion module, a decoupling module, an ambiguous sample elimination module, and a result output module. Among them, the model construction module is used to build a defect detection model and receive the image data to be analyzed; the feature extraction module is used to extract three image features from the image data using the defect detection model; the feature fusion module is used to perform a multi-scale fusion operation on the three image features to obtain fusion features; the decoupling module is used to perform a decoupling operation on the fusion features to obtain regression features, classification features, and regression offset features; the ambiguous sample elimination module is used to construct a quality alignment matrix and a fine-grained compensation index, and use the quality alignment matrix to obtain positive sample points. At the same time, the quality alignment matrix is ​​normalized using the fine-grained compensation coefficient, and a dual-weight loss is constructed. The quality alignment matrix is ​​used as the positive sample weight and ambiguous samples are eliminated. The result output module is used to obtain the final result of defect detection based on the eliminated samples, the classification features, and the corrected regression features, and output it.

[0038] In a third aspect, a multi-category defect detection device is provided, comprising: a processor and a memory storing computer program instructions, wherein the processor reads and executes the computer program instructions to implement a multi-category defect detection method.

[0039] In a fourth aspect, a computer-readable storage medium is proposed, on which computer program instructions are stored. When the computer program instructions are executed by a processor, a multi-category defect detection method is implemented.

[0040] Beneficial effects: The present invention proposes a multi-category defect detection method and system based on drone distribution network overhead lines. By adopting a fine-grained sample allocation method, the quality alignment matrix is ​​calculated, and the coefficients of difficult samples in the quality alignment matrix are compensated by a preset fine-grained compensation index, and the quality alignment matrix is ​​normalized, thereby solving the problem of too low matrix scores at the beginning of training.

[0041] In addition, the present invention also proposes a dual-weight loss, using the quality alignment matrix as the weight of pos_loss to improve the consistency of model classification and regression and improve the ability to mine difficult samples; at the same time, IOU is introduced into neg_loss to reduce the classification score of anchors with lower IOU, thereby reducing the false detection of multi-category defect detection models. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a data processing flow chart of the present invention.

[0043] Figure 2 This is a data processing diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In the following description, numerous specific details are provided to provide a more thorough understanding of the present invention. However, it will be apparent to those skilled in the art that the present invention may be practiced without one or more of these details. In other instances, certain technical features well known in the art are not described to avoid confusion with the present invention.

[0045] Example 1

[0046] In one embodiment, to address the problems in the existing technology such as large computing resources occupied by defect detection models, high false detection rate, and high maintenance cost, a multi-category defect detection method based on drone distribution overhead lines is proposed. By adopting fine-grained sample matching and constructing a dual-weight loss method, the detection performance of the defect detection model is optimized and the accuracy of multi-category defect detection is improved.

[0047] Specifically, such as Figure 1 As shown in FIG, the multi-category defect detection method for overhead lines of a distribution network based on a drone includes the following steps:

[0048] Step 1: Build a defect detection model and receive image data to be analyzed;

[0049] Step 2: extracting three image features from the image data using the defect detection model;

[0050] Step 3: Perform a multi-scale fusion operation on the three image features to obtain fusion features;

[0051] Step 4: Perform decoupling operation on the fused features to obtain regression features, classification features, and regression offset features;

[0052] Step 5: Construct a quality alignment matrix and a fine-grained compensation index, use the quality alignment matrix to obtain positive sample points, and normalize the quality alignment matrix using the fine-grained compensation coefficient.

[0053] Step 6: Construct a dual-weight loss, use the quality alignment matrix as the pos_loss weight, and eliminate ambiguous samples;

[0054] Step 7: Based on the samples after disambiguation, the final result of defect detection is obtained according to the classification features and the corrected regression features.

[0055] Example 2

[0056] In a further embodiment based on the first embodiment, a defect detection model used in the process of performing multi-category defect detection includes: a backbone module, a neck module, a decoupling module, a correction module and an output module.

[0057] For the received image data, the defect detection model first uses the backbone module to perform feature extraction operations to obtain three output features; secondly, the neck module is used to perform multi-scale fusion of the three output features to obtain fusion features, and the fusion features are transmitted to the decoupling module; thirdly, the decoupling module decouples the fusion results to obtain classification features, regression features and regression offset features; from then on, the correction module is used to fuse the regression features and regression offset features to generate regression coordinates, and then obtain the specific coordinates of the detection frame; finally, the output module outputs the corrected regression features and classification features.

[0058] In actual applications, the images collected by the drone distribution network line are large-size images with a pixel size of 5000*3000. To improve the performance of defect detection, the backbone module uses CSPnet and PAN-FPN as the neck module, Decoupled-Head as the decoupling module, and Box Refine as the correction module. The CSPnet backbone module consists of five downsampling modules with a stride of 2; the Decoupled-Head decoupling module consists of three 1×1 convolutions with a stride of 1; the BoxRefine correction module consists of deformable convolution, 1×1 convolution, and a BN layer. During data processing, the regression offset feature is used as the offset matrix of the deformable convolution to offset the input regression feature through interpolation. The parameters of the deformable convolution are set to a fixed constant matrix, and finally the regression feature is output after 1×1 convolution and BN. The Box Refine correction module used in this embodiment can adaptively optimize and correct the regression coordinates. Through the output of three regression offset features, the regression offset features can learn edge information. As the offset matrix of the deformable convolution, it can effectively adaptively correct the regression features and output more accurate regression coordinates. The regression offset features can adaptively offset the input features according to the image information in the deformable convolution, thereby outputting more accurate regression coordinates. Therefore, the proposed Box Refine correction module is friendly to small-scale defects in the distribution network, can stably improve the model regression performance, and increase the defect positioning capability.

[0059] In a preferred embodiment, the image size received by the defect detection model is preset to 2240*2240*3. After passing through the CSPnet backbone module, three output features are obtained, with sizes of (b, 280, 280, 256), (b, 140, 140, 256), and (b, 70, 70, 256). Secondly, the three output features are transmitted to the PAN-FPN neck module for multi-scale fusion to obtain fusion features. Thirdly, the fusion features are transmitted to the Decoupled-Head decoupling module to obtain three classification features (b, 280, 280, c), (b, 140, 140, c ), (b,70,70,c) and three regression features (b,280,280,4), (b,140,140,4), (b,70,70,4) and three regression offset features (b,280,280,8), (b,140,140,8), (b,70,70,8); then, the correction module is used to combine and analyze the regression offset features and the regression features to obtain the corrected regression coordinates, and then the specific coordinates of the detection frame are obtained. The output module outputs the corrected regression coordinates and classification features; finally, the final defect detection result is obtained based on the specific coordinates of the detection frame and the classification features. Where b is the batch size and c is the number of defect categories. In this embodiment, the number of defect categories is 16.

[0060] Example 3

[0061] In a further embodiment based on Example 1, in order to improve the defect recall rate and avoid gradient explosion, when obtaining the specific coordinates of the detection frame, NMS non-maximum suppression is used to filter redundant target frames, and the target frame with the highest classification score is used as the main body to calculate the IOU with other target frames. When the IOU is greater than a certain threshold, it is considered to be a redundant frame and is filtered out.

[0062] To improve the consistency of classification and regression, a quality alignment matrix and a fine-grained compensation index are constructed. The fine-grained compensation index is used to compensate for the coefficients of difficult samples in the quality alignment matrix. The quality alignment matrix is ​​also normalized, thereby addressing the problem of low matrix scores at the beginning of defect detection model training. Simultaneously, a dual-weight loss is constructed, using the quality alignment matrix as a weight for positive samples and eliminating ambiguous samples, thereby improving the consistency of model classification and regression and enhancing the ability to mine difficult samples. Difficult samples are small, infrequent, and prone to deformation and occlusion.

[0063] Among them, the expression of the quality alignment matrix is:

[0064] Align Metric=cls 0.5 *iou 5

[0065] Where cls represents the classification score output by the decoupling module; iou represents the intersection-over-union ratio between the regression coordinates output by the decoupling module and the true target box. Due to the small size and occlusion of some difficult samples in the quality alignment matrix AlignMetric, they cannot obtain high classification scores and iou. Compared with ordinary samples, the metric of difficult samples will be smaller than that of ordinary samples, resulting in a small proportion of difficult samples in the loss loss and inability to converge effectively. Therefore, in order to improve the metric of difficult samples, a fine-grained compensation index C is constructed, and the corresponding expression is:

[0066]

[0067] Where w represents the width of the true target box; h represents the height of the true target box; cls represents the classification score of the defect detection model; and const represents the difficulty coefficient. const is derived from distribution network defect data statistics. In this example, all defect sizes in the training data are counted, sorted from largest to smallest, and the size that accounts for 10% is assigned to const (const = 42 in this example).

[0068] Then, the fine-grained compensation coefficient C is combined with the quality alignment matrix Align Metric to obtain the optimized quality alignment matrix. The corresponding expression is:

[0069] Align Metric′=(Align Metric) C

[0070] The Align Metric′ considers the width and height of each GT real target box in a more fine-grained manner, effectively compensating for the metric score of difficult samples, thereby increasing the proportion of difficult samples in the loss and optimizing gradient return.

[0071] After obtaining the optimized Align Metric′, sort it and select the top 10 positive sample points for each true target box. Since the defect detection model parameters are randomly initialized at the beginning of training, the classification score of the model is low. Therefore, the optimized quality alignment matrix Align Metric′ is normalized. The corresponding expression is:

[0072] Pos_align_metrics=Align Metric′.amax(-1)

[0073] norm_Align_Metric=Align Metric′ / Pos_align_metrics

[0074] In the formula, the size of Align Metric′ is (b, obj_num, grid_num), where objnum represents the number of ground-truth target boxes in the current image, and grid_num represents the number of grids occupied by the output features. In this embodiment, grid_num is 70*70+140*140+280*280; amax(-1) means taking the maximum value in the dimension of grid_num, so the size of Pos_align_metrics is (b, obj_num, 1), which represents the maximum value of the metric quality of each ground-truth target box. Align Metric can be normalized by dividing it by Pos_align_metrics to ensure that the maximum value of the metric for each ground-truth target box is 1.

[0075] In order to improve the recall rate of distribution network defect detection and reduce false detections, this embodiment proposes a dual-weight loss to decouple the negative sample loss from the positive sample loss. As training progresses, the positive samples selected by fine-grained sample matching gradually stabilize, and the classification score output by the model presents a Gaussian distribution. When the IOU is inconsistent with the classification score, it is easy for an object to have multiple detection frames (which NMS cannot filter out), resulting in false detection. In this embodiment, the IOU score of the anchor is used as the evaluation coefficient when designing the negative sample loss, limiting the classification score of the anchor with a low IOU score. The corresponding expression is:

[0076] neg_weight = (1-iou) 2 *cls.sigmoid()

[0077] neg_loss=BCE(cls.sigmoid(), neg_label)*neg_weight

[0078] In the formula, iou represents the intersection-over-union ratio of the regression coordinates output by the model and the true coordinate frame; cls.sigmoid() represents the sigmoid activation of the last dimension of the classification feature; BCE represents cross entropy; neg_label represents all anchor anchor frames.

[0079] To ensure high consistency in model output for classification and regression tasks and improve the recall rate of defect categories, this embodiment uses the normalized quality alignment matrix norm Align Metric as a fine-grained weight in the classification positive sample loss and regression loss respectively. The corresponding expressions are:

[0080] pos_weight = (1-cls.sigmoid) 2 *pos_label

[0081] pos_loss=BCE(cls.sigmoid(), norm_Align_Metric)*pos_weight

[0082] loss = neg_loss + pos_loss

[0083] Where pos_label represents the positive sample anchor box; neg_loss represents the negative sample anchor box.

[0084] Finally, neg_loss and pos_loss are added together to form a dual-weight loss. This dual-weight loss can simultaneously achieve high classification and IOU consistency, as well as the ability to mine difficult samples, effectively reducing false positives and improving defect detection recall. To avoid gradient explosion, this example identifies the indices of ambiguous samples equal to 0 in cls.sigmoid() and sets these indices to zero in the positive sample weights, truncating the gradient.

[0085] Optionally, the score proposed in this example refers to normalizing the meaningless feature value into a score value between 0 and 1 through simoid, and the classification feature is obtained by simoid.

[0086] Optionally, this embodiment uses the quality alignment matrix as the weight for pos_loss positive samples to improve the consistency of model classification and regression and enhance the ability to mine difficult samples. Introducing IOU into neg_loss negative samples lowers the classification score of anchor boxes with lower IOU, reducing model false positives. Furthermore, the indices of ambiguous samples that are equal to 0 in cls.sigmoid() are found and set to zero in the positive sample weights to truncate the gradient and avoid gradient explosion.

[0087] Example 4

[0088] In one embodiment, a multi-category defect detection system is proposed for implementing a multi-category defect detection method. The system includes the following modules: a model construction module, a feature extraction module, a feature fusion module, a decoupling module, an ambiguous sample elimination module, and a result output module. The model construction module is used to construct a defect detection model and receive image data to be analyzed. The feature extraction module is used to extract three image features from the image data using the defect detection model. The feature fusion module is used to perform a multi-scale fusion operation on the three image features to obtain fused features. The decoupling module is used to perform a decoupling operation on the fused features to obtain regression features, classification features, and regression offset features. The ambiguous sample elimination module is used to construct a quality alignment matrix and a fine-grained compensation index, and to obtain positive sample points using the quality alignment matrix. The fine-grained compensation coefficient is used to normalize the quality alignment matrix, construct a dual-weight loss, use the quality alignment matrix as the positive sample weight, and eliminate ambiguous samples. The result output module is used to obtain the final defect detection result based on the disambiguated samples, the classification features, and the corrected regression features, and output it.

[0089] In a further embodiment, a weight loss module is proposed, which satisfies the high consistency of classification and IOU and the ability to mine difficult samples by constructing a dual weight loss method, thereby effectively reducing false detections and improving defect detection recall.

[0090] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes may be made to it in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A multi-category defect detection method for overhead lines based on drone distribution network, characterized in that: The following steps are involved: Step 1: Build a defect detection model and receive image data to be analyzed; Step 2: extracting three image features from the image data using the defect detection model; Step 3: Perform a multi-scale fusion operation on the three image features to obtain fusion features; Step 4: Perform decoupling operation on the fused features to obtain regression features, classification features, and regression offset features; Step 5: Construct a quality alignment matrix and a fine-grained compensation index, use the quality alignment matrix to obtain positive sample points, and normalize the quality alignment matrix using the fine-grained compensation coefficient. Step 6: Construct a dual-weight loss, use the quality alignment matrix as the weight of positive samples, and eliminate ambiguous samples; Step 7: Based on the samples after disambiguation, the final result of defect detection is obtained according to the classification features and the corrected regression features.

2. The multi-category defect detection method based on UAV distribution network overhead lines according to claim 1 is characterized in that: The defect detection model includes: a backbone module, a neck module, a decoupling module, a correction module and an output module. During the data processing process, the backbone module is first used to perform feature extraction operations to obtain three output features; secondly, the neck module is used to perform multi-scale fusion of the three output features to obtain fusion features, and the fusion features are transmitted to the decoupling module; thirdly, the decoupling module decouples the fusion results to obtain classification features, regression features and regression offset features; thirdly, the correction module is used to fuse the regression features and regression offset features to generate regression coordinates, and then the specific coordinates of the detection frame are obtained; finally, the output module outputs the corrected regression features and classification features.

3. The multi-category defect detection method based on UAV distribution network overhead lines according to claim 1 is characterized in that: During the process of performing classification and regression in the defect detection model, a quality alignment matrix is ​​constructed and optimized using a fine-grained compensation coefficient, thereby performing fine-grained sample matching.

4. The multi-category defect detection method based on UAV distribution network overhead lines according to claim 3 is characterized in that: The expression of the quality alignment matrix is: Align Metric=cls 0.5 *iou 5 Where cls represents the classification score output by the decoupling module; iou represents the intersection-over-union ratio between the regression coordinates output by the decoupling module and the true target frame; The expression of the fine-grained compensation coefficient is: Where w represents the width of the true target frame; h represents the height of the true target frame; cls represents the classification score of the defect detection model; const represents the difficulty coefficient; The expression for combining the quality alignment matrix with the fine-grained compensation coefficient is: Align Metric′=(Align Metric) C After obtaining the Align Metric′, a sorting operation is performed and the top 10 points are selected as positive sample points for each true target box.

5. The multi-category defect detection method based on drone distribution network overhead lines according to claim 4 is characterized in that: After obtaining the positive sample points, the normalization operation is performed, and the corresponding expression is: Pos_align_metrics=Align Metric′.amax(-1) norm_Align_Metric=Align Metric′ / Pos_align_metrics In the formula, the size of Align Metric′ is (b, obj_num, grid_num), where objnum represents the number of G real target boxes in the current image, grid_num represents the number of grids occupied by output features, and b represents the batch size; amax(-1) represents the maximum value in the dimension of grid_num; Pos_align_metrics represents the maximum value of each GT real target box in the metric quality.

6. The multi-category defect detection method based on UAV distribution network overhead lines according to claim 1 is characterized in that: By constructing a dual-weight loss, the negative sample loss is decoupled from the positive sample loss; When constructing the negative sample loss, the IOU score of the anchor box is used as the evaluation coefficient, and the corresponding expression is: neg_weight=(1-iou) 2 *cls.sigmoid() neg_loss=BCE(cls.sigmoid(), neg_label)*neg_weight In the formula, iou represents the intersection-over-union ratio of the regression coordinates output by the model and the true coordinate frame; cls.sigmoid() represents the sigmoid activation of the last dimension of the classification feature; BCE represents cross entropy; neg_label represents all anchor anchor frames.

7. The multi-category defect detection method based on UAV distribution network overhead lines according to claim 6 is characterized in that: The obtained normalized quality alignment matrix is ​​used as the fine-grained weight for classifying positive samples and regression loss. The corresponding expression is: pos_weight=(1-cls.sigmoid) 2 *pos_label pos_loss=BCE(cls.sigmoid(), norm_Align_Metric)*pos_weight loss = neg_loss + pos_loss Where pos_label represents the positive sample anchor box; neg_loss represents the negative sample anchor box.

8. A multi-category defect detection system based on drone distribution network overhead lines, used to implement the multi-category defect detection method according to any one of claims 1 to 7, characterized in that: Includes the following modules: a model building module configured to build a defect detection model and receive image data to be analyzed; a feature extraction module configured to extract three image features from the image data using the defect detection model; A feature fusion module is configured to perform a multi-scale fusion operation on the three image features to obtain a fused feature; The decoupling module is configured to perform a decoupling operation on the fused features to obtain regression features, classification features, and regression offset features; The ambiguous sample elimination module is configured to construct a quality alignment matrix and a fine-grained compensation index, and use the quality alignment matrix to obtain positive sample points. The quality alignment matrix is ​​normalized using the fine-grained compensation coefficient, and a dual-weight loss is constructed. The quality alignment matrix is ​​used as the positive sample weight to eliminate ambiguous samples. The result output module is set to obtain the final result of defect detection based on the samples after disambiguation, classification features and corrected regression features, and output it.

9. A multi-category defect detection device, characterized in that: The device comprises: a processor and a memory storing computer program instructions; The processor reads and executes the computer program instructions to implement the multi-category defect detection method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the multi-category defect detection method according to any one of claims 1 to 7.

Images