Auto parts defect identification method, system, computer equipment and storage medium

Abstract

The present invention discloses a method, system, computer equipment and storage medium for identifying defects in automotive parts. A high-frequency structural feature image is obtained by performing discrete Fourier transform on an original image. In order to solve the problem of limited training samples in industrial defect detection, the feature expression of the target is enriched by fusing the high-frequency structural features of the target in the original image into the original features of the defect. This enables a deep learning model to fully learn the feature expression of the defect target with fewer training samples, thereby improving the classification performance of known categories. The obtained defect feature vector is used to predict the known class category through a fully connected classification network in an automotive part defect recognition method based on high-frequency structural feature fusion enhancement and class mutual information constraint, and to judge whether the input sample data belongs to an unknown class, ultimately achieving accurate defect detection and unknown defect recognition in industrial defect detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial defect detection, and in particular relates to an automobile parts defect recognition method, system, computer equipment and storage medium. Background Art

[0002] In modern manufacturing, automotive component defect detection is not just a component of quality assurance; it's a critical component in enhancing a company's competitiveness. An efficient and accurate defect detection system ensures product consistency and reliability, thereby meeting customer demand for high-quality products and enhancing a company's market reputation. Furthermore, early defect detection and classification enable companies to quickly implement corrective actions, reducing scrap and rework costs, and improving overall production line efficiency.

[0003] However, faced with the increasingly complex structures of automotive parts and the diverse types of defects, traditional defect detection methods are no longer sufficient. Manual inspection is not only inefficient and susceptible to operator subjective factors, but also difficult to adapt to high-volume, high-speed production environments. Therefore, automation and intelligentization have become the inevitable trend in the development of defect detection technology.

[0004] In recent years, artificial intelligence technologies such as machine learning and deep learning have made significant progress in automotive parts defect detection. These technologies can autonomously learn defect characteristics using large amounts of training data, enabling automatic recognition and classification of complex defect patterns. For example, convolutional neural networks (CNNs) excel in image recognition and have been widely used in industrial defect detection, significantly improving detection accuracy and speed.

[0005] Despite this, current automated inspection systems still face several challenges. First, the need for large datasets remains a bottleneck. Achieving high-precision inspection without relying on massive samples remains a pressing issue. Second, the lack of diversity and representativeness of samples can lead to poor model generalization, making it difficult to address the various unseen defects that may arise during production. Furthermore, subtle differences between different defect types can lead to misjudgments, compromising detection accuracy.

[0006] To this end, the present invention proposes a method, system, computer device and storage medium for automobile parts defect recognition based on high-frequency structural feature fusion enhancement and quasi-mutual information constraint. Summary of the Invention

[0007] In response to the above technical problems, the present invention provides a method, system, computer device and storage medium for identifying defects in automotive parts.

[0008] The technical solution adopted by the present invention to solve the technical problem is:

[0009] A method for identifying defects in automobile parts, comprising the following steps:

[0010] S100: Acquire an original image containing defects of automobile parts, and perform discrete Fourier transform on the original image to obtain a high-frequency structural feature image;

[0011] S200: Build a defect recognition network, including a feature extraction network, a feature fusion module, and a classification network;

[0012] S300: Inputting the high-frequency structural feature image and the original image into the feature extraction network for feature extraction, thereby obtaining the original defect features and the defect structural features; inputting the original defect features and the defect structural features into the feature fusion module for fusion, thereby obtaining the defect features after the high-frequency structural features are enhanced and fused;

[0013] S400: Calculating the mutual information between different categories of defect features based on the defect features enhanced and fused with high-frequency structural features to obtain mutual information loss;

[0014] S500: Inputting the defect features after enhancing and fusing the high-frequency structural features into the classification network, combining the preset loss function to obtain the known class classification loss, and combining the known class classification loss and the mutual information loss to constrain the defect recognition network training. When the preset training end condition is met, a trained defect recognition network is obtained;

[0015] S600: Acquire a real-time input image, input the real-time input image into a trained defect recognition network, make predictions of known class categories, and determine whether the input sample data belongs to an unknown class, thereby completing the defect recognition of automotive parts.

[0016] Preferably, S100 specifically includes:

[0017] ;

[0018] ;

[0019] in is the high-frequency structural feature image, is the input image, is the discrete Fourier transform, is the inverse discrete Fourier transform, is the high frequency component threshold, is the discrete Fourier transform of the input original image, N is the total number of pixels in the image, is the position coordinate of each pixel in the image, are the position coordinates of each pixel in the image after Fourier transformation.

[0020] Preferably, S300 specifically includes:

[0021] ;

[0022] ;

[0023] ;

[0024] in is the feature extraction network, is the input image, is the original feature of the defect extracted, is the high-frequency structural feature image, is the extracted defect structure feature, It is the defect feature after the high-frequency structural features are enhanced and fused. is a convolution operation with a kernel size of 1, It is a feature concatenation operation.

[0025] Preferably, the feature extraction network consists of three layers of convolution and pooling operations, and the convolution and pooling operations of each layer are specifically:

[0026] ;

[0027] in is the input feature map or the output of the previous layer, It is a convolution operation with a convolution kernel of 3. is the maximum pooling with a scale of 2, is the batch normalization operation, is the output feature map.

[0028] Preferably, S400 specifically includes:

[0029] ;

[0030] in is the feature of the i-th known category, K is the number of training samples of the k-th category, is the variable x and y and The joint probability distribution of represents the mutual information between feature vectors of different defect categories, represents the mutual information loss.

[0031] Preferably, in S500, the known class classification loss and the mutual information loss are combined to constrain the defect recognition network training, specifically:

[0032] ;

[0033] ;

[0034] in, is the total training loss of the automotive parts defect recognition method with high-frequency structural feature fusion enhancement and mutual information constraint, is the known class classification loss, is the mutual information loss between different categories of defect features, N is the number of samples, C is the number of categories, is the true label, is the predicted label, is the natural logarithm.

[0035] The automotive parts defect recognition system includes a high-frequency structural feature image acquisition module, a defect recognition network building module, a feature extraction and fusion module, a mutual information loss calculation module, a classification and defect recognition network training module, and an automotive parts defect recognition module.

[0036] A high-frequency structural feature image acquisition module is used to acquire an original image containing defects in automotive parts and perform discrete Fourier transform on the original image to obtain a high-frequency structural feature image;

[0037] Defect recognition network building module, used to build a defect recognition network, including feature extraction network, feature fusion module and classification network;

[0038] The feature extraction and fusion module is used to receive the high-frequency structural feature image and the original image to perform feature extraction to obtain the original defect features and the defect structural features; the original defect features and the defect structural features are fused to obtain the defect features after the high-frequency structural features are enhanced and fused;

[0039] A mutual information loss calculation module is used to calculate the mutual information between different categories of defect features based on the defect features enhanced and fused with high-frequency structural features to obtain the mutual information loss;

[0040] The classification and defect recognition network training module is used to input the defect features after enhancing and fusing high-frequency structural features into the classification network, combine the preset loss function to obtain the known class classification loss, and jointly constrain the defect recognition network training with the known class classification loss and mutual information loss. When the preset training end condition is met, a trained defect recognition network is obtained;

[0041] The automotive parts defect recognition module is used to obtain real-time input images, input the real-time input images into the trained defect recognition network, predict the known class category, and determine whether the input sample data belongs to the unknown class to complete the automotive parts defect recognition.

[0042] A computer device includes a memory and a processor, wherein the memory stores a computer program and the processor implements the steps of an automobile parts defect identification method when executing the computer program.

[0043] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for identifying defects in automobile parts.

[0044] The above-mentioned automotive parts defect recognition method, system, computer equipment and storage medium first obtain the high-frequency structural feature image of the defect image based on discrete Fourier transform, and then the original image and the high-frequency structural feature image are jointly input into the feature extraction network for feature extraction, and feature fusion enhancement is performed to enrich the feature expression of the defect target; finally, the fused defect feature map is used to calculate the mutual information between different categories of defect features through a class mutual information constraint strategy, and the classification loss is combined as the training loss to constrain the network training, thereby enhancing the distinguishability between different categories of defect features and improving detection accuracy.

[0045] After obtaining the trained network, the high-frequency structural feature image of the input defect image is first obtained based on the discrete Fourier transform. The original image and the high-frequency structural feature image are input into the feature extraction network for feature extraction, and then input into the feature fusion module for feature fusion enhancement. The fused feature vector is then classified into known categories and discriminated against unknown categories through a classifier. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Flowchart of a method for identifying defects in automotive parts according to one embodiment of the present invention;

[0047] Figure 2 Schematic diagram of the principle of a method for identifying defects in automotive parts according to one embodiment of the present invention;

[0048] Figure 3 This is a structural diagram of a feature extraction network in one embodiment of the present invention;

[0049] Figure 4 This is a structural diagram of a feature fusion module in one embodiment of the present invention;

[0050] Figure 5 3 is a comparison chart of test results of the existing method and the automobile parts defect recognition method of the present invention; wherein, (a) is the feature distribution of different sample features extracted by the existing method, and (b) is the feature distribution of different sample features extracted by the automobile parts defect recognition method of the present invention. DETAILED DESCRIPTION

[0051] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below with reference to the accompanying drawings.

[0052] In one embodiment, Figure 1 and Figure 2As shown, a method for identifying defects in automobile parts comprises the following steps:

[0053] S100: Acquire an original image containing defects of automobile parts, and perform discrete Fourier transform on the original image to obtain a high-frequency structural feature image;

[0054] S200: Build a defect recognition network, including a feature extraction network, a feature fusion module, and a classification network;

[0055] S300: Inputting the high-frequency structural feature image and the original image into the feature extraction network for feature extraction, thereby obtaining the original defect features and the defect structural features; inputting the original defect features and the defect structural features into the feature fusion module for fusion, thereby obtaining the defect features after the high-frequency structural features are enhanced and fused;

[0056] S400: Calculating the mutual information between different categories of defect features based on the defect features enhanced and fused with high-frequency structural features to obtain mutual information loss;

[0057] S500: Inputting the defect features after enhancing and fusing the high-frequency structural features into the classification network, combining the preset loss function to obtain the known class classification loss, and combining the known class classification loss and the mutual information loss to constrain the defect recognition network training. When the preset training end condition is met, a trained defect recognition network is obtained;

[0058] S600: Acquire a real-time input image, input the real-time input image into a trained defect recognition network, make predictions of known class categories, and determine whether the input sample data belongs to an unknown class, thereby completing the defect recognition of automotive parts.

[0059] Specifically, this paper proposes a defect recognition method for automotive parts based on high-frequency structural feature fusion enhancement and class mutual information constraints. This method can quickly adapt to complex component characteristics, efficiently identify and classify various defects, and accurately detect unknown defects. This method is crucial for improving production efficiency, reducing costs, and ensuring product quality. The core concept of this method is to enrich the target feature representation by integrating the high-frequency structural features of defects, addressing the problem of limited training samples. Furthermore, inter-class feature mutual information constraints are used to increase the discriminative margin between defect features of different classes, addressing the difficulty in distinguishing between known and unknown classes. Specifically, this method proposes a high-frequency structural feature fusion enhancement strategy to meticulously extract and fuse the high-frequency structural features of defects. This enables the network to comprehensively learn the feature representation of defect targets even under constrained training sample conditions, thereby extracting robust features for known classes. This enhancement improves the distinguishability between known and unknown classes, facilitating accurate identification of unknown classes. Furthermore, a class mutual information constraint strategy is proposed in the automotive parts defect recognition method based on high-frequency structural feature fusion enhancement and class mutual information constraint to measure and mitigate the mutual information between defect features of different categories. By doing so, the defect features of different known classes are ensured to be independent, thereby improving their distinguishability and significantly improving the recognition performance of both known and unknown classes.

[0060] In one embodiment, S100 specifically includes:

[0061] ;

[0062] ;

[0063] in is the high-frequency structural feature image, is the input image, is the discrete Fourier transform, is the inverse discrete Fourier transform, is the high frequency component threshold, is the discrete Fourier transform of the input original image, N is the total number of pixels in the image, is the position coordinate of each pixel in the image, is the position coordinate of each pixel in the Fourier transformed image. Furthermore, since defects generally belong to high-frequency components and background belongs to low-frequency components, the high-frequency component threshold is used to eliminate the low-frequency background components and only retain the defect part.

[0064] Specifically, in this example, the original input image size is 256×256. Limited training samples are a major factor affecting industrial defect detection performance. This is because a small number of training samples prevents the detection network model from effectively learning the characteristic representations of defect targets, thus failing to accurately detect and classify defect categories. Furthermore, some low-frequency defects cannot be effectively detected due to the inability to collect training data during the data collection phase.

[0065] To address these issues, the present invention enriches the target's feature representation by fusing the high-frequency structural features of the target in the original image with the original features of the defect. This enables the deep learning model to fully learn the feature representation of the defect target with fewer training samples, thereby improving the classification performance of known categories.

[0066] In one embodiment, Figure 3 and Figure 4 As shown, S300 is specifically:

[0067] ;

[0068] ;

[0069] ;

[0070] in is the feature extraction network, is the input image, is the original feature of the defect extracted, is the high-frequency structural feature image, is the extracted defect structure feature, It is the defect feature after the high-frequency structural features are enhanced and fused. is a convolution operation with a kernel size of 1, It is a feature concatenation operation.

[0071] In one embodiment, Figure 3 As shown in the figure, the feature extraction network consists of three layers of convolution and pooling operations. The convolution and pooling operations of each layer are as follows:

[0072] ;

[0073] in is the input feature map or the output of the previous layer, It is a convolution operation with a convolution kernel of 3. is the maximum pooling with a scale of 2, is the batch normalization operation, is the output feature map.

[0074] Specifically, the high-frequency structural feature image of the feature extraction network and the high-frequency structural features in the original features With the original features Will be extracted and eventually fused into high-frequency structural feature enhancement and fused defect features To address the limited number of training samples in industrial defect detection, this paper enriches the target's feature representation by fusing the high-frequency structural features of the target in the original image with the original features of the defect. This enables the deep learning model to fully learn the feature representation of the defect target with fewer training samples, thereby improving the classification performance of known categories.

[0075] In one embodiment, S400 specifically includes:

[0076] ;

[0077] in is the feature of the i-th known category, K is the number of training samples of the k-th category, is the variable x and y and The joint probability distribution of represents the mutual information between feature vectors of different defect categories, represents the mutual information loss.

[0078] Specifically, to reduce the differences between defect features of different types, each feature's contribution to the classification task is made more independent and significant, thereby improving classifier performance. Mutual information measures the correlation between two variables. If the mutual information between defect features of different categories is high, it indicates a strong dependency between these features, leading to feature redundancy. By reducing mutual information, redundancy between features can be reduced, making each feature's contribution to the classification task more independent and significant, thereby improving classifier performance. When mutual information between features is low, each feature contributes more independently to the classification task, better capturing the differences between different categories and thus enhancing the feature's discriminative power.

[0079] By calculating the mutual information between the feature vectors of different defect categories and use it as loss , and by minimizing the mutual information between the prototype vectors of different categories of defect features To increase the difference between features of different categories, thus accurately classifying known categories in open set recognition of industrial defects.

[0080] In one embodiment, in S500, the known class classification loss and the mutual information loss are combined to constrain the defect recognition network training, specifically:

[0081] ;

[0082] ;

[0083] in, is the total training loss of the automotive parts defect recognition method with high-frequency structural feature fusion enhancement and mutual information constraint, is the known class classification loss, is the mutual information loss between different categories of defect features, N is the number of samples, C is the number of categories, is the true label, is the predicted label, is the natural logarithm.

[0084] Specifically, the classification loss generated by the classification network is Mutual information loss between different categories of defect features , to achieve efficient training of automobile parts defect recognition methods based on high-frequency structural feature fusion enhancement and class mutual information constraints, so that the network model can effectively classify known classes and identify unknown classes.

[0085] Furthermore, for the test image, the high-frequency structural image is obtained using the same method as in the training phase, and the defect feature vector after structural feature fusion enhancement is obtained. Finally, the defect feature vector is classified into known classes and identified into unknown classes through the classification network.

[0086] The above-mentioned automotive parts defect recognition method proposes a high-frequency structural feature fusion enhancement strategy and a class mutual information constraint strategy, respectively, which solves the problem of inability to accurately identify known and unknown defects in industrial open set recognition due to the small difference between restricted training samples and defect categories, and significantly improves the performance of industrial defect open set recognition. Figure 5 As shown in the figure, the automotive parts defect recognition method based on high-frequency structural feature fusion enhancement and class mutual information constraint can effectively distinguish features of known and unknown classes. Experiments show that the accuracy rate increased by 10%, the F1 score increased by 22%, and the area under the receiver operating characteristic curve increased by 10%.

[0087] The automotive parts defect recognition system includes a high-frequency structural feature image acquisition module, a defect recognition network building module, a feature extraction and fusion module, a mutual information loss calculation module, a classification and defect recognition network training module, and an automotive parts defect recognition module.

[0088] A high-frequency structural feature image acquisition module is used to acquire an original image containing defects in automotive parts and perform discrete Fourier transform on the original image to obtain a high-frequency structural feature image;

[0089] Defect recognition network building module, used to build a defect recognition network, including feature extraction network, feature fusion module and classification network;

[0090] The feature extraction and fusion module is used to receive the high-frequency structural feature image and the original image to perform feature extraction to obtain the original defect features and the defect structural features; the original defect features and the defect structural features are fused to obtain the defect features after the high-frequency structural features are enhanced and fused;

[0091] A mutual information loss calculation module is used to calculate the mutual information between different categories of defect features based on the defect features enhanced and fused with high-frequency structural features to obtain the mutual information loss;

[0092] The classification and defect recognition network training module is used to input the defect features after enhancing and fusing high-frequency structural features into the classification network, combine the preset loss function to obtain the known class classification loss, and jointly constrain the defect recognition network training with the known class classification loss and mutual information loss. When the preset training end condition is met, a trained defect recognition network is obtained;

[0093] The automotive parts defect recognition module is used to obtain real-time input images, input the real-time input images into the trained defect recognition network, predict the known class category, and determine whether the input sample data belongs to the unknown class to complete the automotive parts defect recognition.

[0094] The specific definitions of the automotive parts defect identification system can be found in the definitions of the automotive parts defect identification method described above and will not be repeated here. Each module in the aforementioned automotive parts defect identification system can be implemented in whole or in part through software, hardware, or a combination thereof. Each of these modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0095] A computer device includes a memory and a processor, wherein the memory stores a computer program and the processor implements the steps of an automobile parts defect identification method when executing the computer program.

[0096] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for identifying defects in automobile parts.

[0097] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media used in the embodiments provided herein may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0098] The above describes in detail the automotive parts defect identification method, system, computer device, and storage medium provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only intended to help understand the core concept of the present invention. It should be noted that those skilled in the art can make various improvements and modifications to the present invention without departing from the principles of the present invention, and such improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A method for identifying defects in automobile parts, characterized in that: The method comprises the following steps: S100: Acquire an original image containing defects of automobile parts, and perform discrete Fourier transform on the original image to obtain a high-frequency structural feature image; S200: Build a defect recognition network, including a feature extraction network, a feature fusion module, and a classification network; S300: Inputting the high-frequency structural feature image and the original image into the feature extraction network for feature extraction, thereby obtaining the original defect features and the defect structural features; inputting the original defect features and the defect structural features into the feature fusion module for fusion, thereby obtaining the defect features after the high-frequency structural features are enhanced and fused; S400: Calculate the mutual information between different types of defect features based on the defect features enhanced and fused with high-frequency structural features to obtain the mutual information loss. S400 specifically includes: ； in is the feature of the i-th known category, K is the number of training samples of the k-th category, is the variable x and y and The joint probability distribution of represents the mutual information between feature vectors of different defect categories, represents the mutual information loss; S500: Inputting the defect features after enhancing and fusing the high-frequency structural features into the classification network, combining the preset loss function to obtain the known class classification loss, and combining the known class classification loss and the mutual information loss to constrain the defect recognition network training. When the preset training end condition is met, a trained defect recognition network is obtained; S600: Acquire a real-time input image, input the real-time input image into a trained defect recognition network, make predictions of known class categories, and determine whether the input sample data belongs to an unknown class, thereby completing the defect recognition of automotive parts.

2. The method according to claim 1, characterized in that S100 is specifically: ； ； in is the high-frequency structural feature image, is the input image, is the discrete Fourier transform, is the inverse discrete Fourier transform, is the high frequency component threshold, is the discrete Fourier transform of the input original image, N is the total number of pixels in the image, is the position coordinate of each pixel in the image, are the position coordinates of each pixel in the image after Fourier transformation.

3. The method according to claim 2, characterized in that S300 is specifically: ； ； ； in is the feature extraction network, is the input image, is the original feature of the defect extracted, is the high-frequency structural feature image, is the extracted defect structure feature, It is the defect feature after the high-frequency structural features are enhanced and fused. is a convolution operation with a kernel size of 1, It is a feature concatenation operation.

4. The method according to claim 3, characterized in that The feature extraction network consists of three layers of convolution and pooling operations. The convolution and pooling operations of each layer are as follows: ； in is the input feature map or the output of the previous layer, It is a convolution operation with a convolution kernel of 3. is the maximum pooling with a scale of 2, is the batch normalization operation, is the output feature map.

5. The method according to claim 4, characterized in that In S500, the known class classification loss and mutual information loss are combined to constrain the defect recognition network training. Specifically: ； ； in, is the total training loss of the automotive parts defect recognition method with high-frequency structural feature fusion enhancement and mutual information constraint, is the known class classification loss, is the mutual information loss between different categories of defect features, N is the number of samples, C is the number of categories, is the true label, is the predicted label, is the natural logarithm.

6. An automobile parts defect recognition system based on the method according to any one of claims 1 to 5, characterized in that: It includes high-frequency structural feature image acquisition module, defect recognition network construction module, feature extraction and fusion module, mutual information loss calculation module, classification and defect recognition network training module and automobile parts defect recognition module. A high-frequency structural feature image acquisition module is used to acquire an original image containing defects in automotive parts and perform discrete Fourier transform on the original image to obtain a high-frequency structural feature image; Defect recognition network building module, used to build a defect recognition network, including feature extraction network, feature fusion module and classification network; The feature extraction and fusion module is used to receive the high-frequency structural feature image and the original image to perform feature extraction to obtain the original defect features and defect structural features; The original defect features and the defect structural features are fused to obtain defect features after high-frequency structural feature enhancement and fusion; A mutual information loss calculation module is used to calculate the mutual information between different categories of defect features based on the defect features enhanced and fused with high-frequency structural features to obtain the mutual information loss; The classification and defect recognition network training module is used to input the defect features after enhancing and fusing high-frequency structural features into the classification network, combine the preset loss function to obtain the known class classification loss, and jointly constrain the defect recognition network training with the known class classification loss and mutual information loss. When the preset training end condition is met, a trained defect recognition network is obtained; The automotive parts defect recognition module is used to obtain real-time input images, input the real-time input images into the trained defect recognition network, predict the known class category, and determine whether the input sample data belongs to the unknown class to complete the automotive parts defect recognition.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images