Industrial defect detection method, system and equipment based on large kernel convolution and feature-level super-resolution
By adopting large-core convolution and feature-level super-resolution technologies in industrial defect detection, the existing methods have solved the problem of insufficient receptive field in capturing complex and small-size defects, and achieved higher detection accuracy and recall.
Patent Information
- Application Number
- CN202510077559.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-17
- Publication Date
- 2025-05-16
AI Technical Summary
The existing industrial defect detection methods have insufficient effective receptive fields in capturing complex defect patterns and fine texture details, resulting in the inability to effectively deal with complex and small-size defects.
The industrial defect detection method based on large core convolution and feature-level super-resolution is adopted to enhance the extraction of global and local context information through feature extraction backbone modules and feature-level super-resolution branch modules, and the full fusion of feature information is achieved through multi-scale feature fusion modules.
It effectively improves the detection accuracy and recall of complex and small-size defects, reduces missed and missed detection, and does not require high computing resources, and has strong versatility.
Smart Images
Figure CN120013887A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of target detection in computer vision, and relates to an industrial defect detection method, system and equipment based on large kernel convolution and feature level super resolution. Background Art
[0002] In industrial manufacturing, defects including scratches, patches, etc. not only pose a threat to customer safety, but also damage the company's reputation. Therefore, in modern intelligent manufacturing, industrial defect detection is crucial for product quality control. Given the presence of multiple objects and real-time monitoring requirements on the production line, the detector must be able to detect complex and small defects efficiently and accurately. Lightweight real-time detection methods such as the YOLO series have been widely adopted due to their suitability for these requirements. However, these methods usually adopt smaller convolution kernels and shallower network structures, resulting in insufficient effective receptive fields to capture complex defect patterns and fine texture details, which are crucial for identifying complex and small-sized defects. We attribute the inability to handle a variety of defects to the lack of sufficient global and local context information extraction in high-level features. To address this problem, many existing methods for industrial defect detection have been proposed from the perspective of global and local context information.
[0003] To enhance the extraction of global context information, many studies have explored attention mechanisms and large kernel convolutions. When dealing with complex spatial structures, attention mechanisms aggregate relevant global context information and reduce unnecessary background information. At the same time, large kernel convolutions recognize complex features and patterns by integrating a wider range of input images. However, high computational complexity and huge resource requirements hinder their further application, thus weakening YOLO's advantage in real-time detection.
[0004] Local contextual information plays an indispensable role in detecting small-sized defects. As a preprocessing step, some object detection tasks utilize the output of a super-resolution network as the input or supervision signal of the detection network to obtain better detection results. However, naive super-resolution solutions greatly increase the computational cost, and the independent training of the super-resolution network and the detection network also limits the ability of super-resolution to improve detection performance. Recently, some studies have proposed a super-resolution auxiliary branch to guide the detector to learn small objects and achieved impressive performance improvements. Specifically, they leveraged an off-the-shelf super-resolution method EDSR to transfer the relevant learning of the spatial dimension to the main feature extraction backbone. However, as an auxiliary branch, this traditional super-resolution method is prone to shift the training focus of lightweight detectors due to its complex network structure, resulting in unsatisfactory detection results and redundant allocation of computational resources. Summary of the invention
[0005] In view of the problems existing in the prior art, the present invention provides an industrial defect detection method and system based on large kernel convolution and feature-level super-resolution, thereby absorbing the advantages of large kernel and super-resolution. The cascaded deep convolution structure obtains a larger effective receptive field through fewer stacked layers, and the super-resolution branch enhances the detailed texture information in higher-level features by reconstructing the fused feature map into the original input image.
[0006] The present invention is achieved through the following technical solutions:
[0007] An industrial defect detection method based on large kernel convolution and feature-level super-resolution, comprising:
[0008] Constructing a defect detection network model, the defect detection network model includes a feature extraction backbone module, a feature-level super-resolution branch module, a multi-scale feature fusion module and a detection module; a large-kernel convolution feature extraction module is introduced into the feature extraction backbone module;
[0009] Obtain images of various defect types in industrial scenes, perform data preprocessing on the images, and obtain training sets and test sets;
[0010] The defect detection network model is trained based on the training set, and the test set is input into the trained defect detection network model for testing to obtain the final defect detection network model;
[0011] The industrial image to be inspected is input into the final defect detection network model to detect defects in the industrial image.
[0012] Preferably, the feature extraction backbone module includes four convolution feature extraction stages, wherein the first stage uses conventional convolution operations to extract features and downsamples the image twice by a factor of 2; the second, third and fourth stages introduce large-kernel convolution feature extraction modules to enhance the extraction of complex defect features, and downsample the image features once by a factor of 2 respectively.
[0013] Preferably, two channel interaction modules consisting of conventional convolution with a kernel of 1*1 and a ReLu activation function are added to the large-kernel convolution feature extraction module.
[0014] Preferably, the feature-level super-resolution branch module fuses the mid-level features and the high-level features through a multi-scale feature fusion module, and then super-resolutions the fused features so that the super-resolution image continuously approaches the original image, thereby introducing the detail information in the original image into the mid-level and high-level features.
[0015] Preferably, the multi-scale feature fusion module uses a feature pyramid structure to output feature maps of different scales.
[0016] Preferably, data preprocessing is performed on the image, and the specific process is as follows:
[0017] Each image is scaled to 1024 pixels on the short side while maintaining its aspect ratio. Then, it is randomly scaled left and right and up and down by 50% and the manually labeled results are inverted accordingly. Then, the image pixels are normalized from 0-255 to 0-1. Finally, all images are padded with peripheral pixels to make the image size 1024*1024 pixels. The image dataset is obtained and then randomly divided into training and test sets.
[0018] Preferably, the defect detection network model is trained based on the training set, and the specific process is:
[0019] The images in the training set are input into the defect detection network model, and are first processed by the feature extraction backbone module to obtain 4 layers of effective feature maps. The 4 layers of effective feature maps are synchronously trained through the feature-level super-resolution branch module to enhance the detail information of small targets of higher-level features. Then, the multi-scale feature fusion module is used to fuse the information and output the 3-layer fused feature maps. Finally, the detection module outputs the defect location and classification results. The total loss function is calculated based on the defect location and classification results and the results of manual calibration. Back propagation is performed through the total loss function, and the Adam optimizer is used to optimize the parameters in the defect detection network model according to the gradient information obtained by back propagation, thereby obtaining the trained defect detection network model.
[0020] Preferably, the specific calculation formula of the total loss function is:
[0021] L total =L de +αL sr
[0022] Where, L total is the total loss function, L sr is the super-resolution loss, α is the super-resolution loss L sr The weight coefficient, L de To detect losses;
[0023] Among them, the super-resolution loss L sr It is obtained by calculating the loss value L1 between the input image I and the super-resolution result SR. The specific formula is:
[0024] L sr =||SR-I|| 1 .
[0025] An industrial defect detection system based on large kernel convolution and feature-level super-resolution, comprising:
[0026] The data preprocessing module is used to obtain images of various defect types in industrial scenes, and obtain training sets and test sets after data preprocessing of the images;
[0027] A model building module, used to build a defect detection network model, wherein the defect detection network model includes a feature extraction backbone module, a feature-level super-resolution branch module, a multi-scale feature fusion module and a detection module;
[0028] The model training module is used to train the defect detection network model with the training set, input the test set into the trained defect detection network model, perform the test, and obtain the final defect detection network model;
[0029] The industrial image detection module is used to input the industrial image to be detected into the final defect detection network model to detect defects in the industrial image.
[0030] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of an industrial defect detection method based on large kernel convolution and feature-level super-resolution are implemented.
[0031] Compared with the prior art, the present invention has the following beneficial technical effects:
[0032] The present invention discloses an industrial defect detection method based on large kernel convolution and feature-level super-resolution. The method comprises a feature extraction trunk module, a feature-level super-resolution branch module, a multi-scale feature fusion module and a detection module. The feature-level super-resolution branch module reintroduces detail texture information into middle and high-level features by super-resolving the fused features, thereby enhancing local context information to detect small-sized defects; the multi-scale feature fusion module realizes full fusion of feature information; the detection module obtains a rectangular anchor frame according to the multi-scale feature map and gives a prediction result. The present invention applies large kernel convolution super-resolution technology to the field of industrial detection, and newly proposes a feature extraction module based on large kernel convolution for sensing the location and feature information of complex defects. At the same time, the feature-level super-resolution branch module reintroduces detail texture information into middle and high-level features to improve the detection accuracy of small-sized defects. Based on this, a defect detection network model is constructed to solve the task of industrial complex and small-sized defect detection. The industrial defect detection method of the present invention fully considers large kernel convolution and super-resolution technology to solve the problem that the existing technology does not sufficiently extract the features of complex and small-sized defects, and can effectively improve the accuracy and recall rate of complex and small-sized defect detection, and reduce missed detection and false detection.
[0033] The large-core convolution feature extraction module and the parallel feature-level super-resolution branch module proposed in the present invention are both based on basic convolution and matrix operations, with consistent input and output data forms, without excessive reliance, and can be conveniently applied to various real-time defect detection models, with broad application prospects. This method introduces large-core convolution and super-resolution technology into the field of industrial inspection, which can effectively improve the detection accuracy and recall rate of complex and small-sized defects, and each module is simple to implement, does not require high computing resources, and has strong versatility. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 This is the structure of the industrial defect detection method based on large kernel convolution and feature level super resolution of the present invention.
[0035] Figure 2 This is the structure of the real-time target detection network in the embodiment of the present invention.
[0036] Figure 3 This is the structure of the large kernel convolution feature extraction module in the embodiment of the present invention.
[0037] Figure 4 This is the structure of the feature-level super-resolution branch in the embodiment of the present invention.
[0038] Figure 5 2 is a diagram of the detection results of an embodiment of the present invention. DETAILED DESCRIPTION
[0039] The present invention is further described in detail below in conjunction with specific embodiments, which are intended to explain the present invention rather than to limit it.
[0040] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not 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 should fall within the scope of protection of the present invention.
[0041] An industrial defect detection method based on large kernel convolution and feature-level super-resolution can simultaneously detect complex and small target defect data sets, and the detection object is an RGB image containing various defect types taken in an industrial scene; the defect detection method comprises four parts: a feature extraction backbone module, a feature-level super-resolution branch module, a multi-scale feature fusion module and a detection module. The feature extraction backbone module is used to extract image features, wherein the first stage of the feature extraction backbone module uses ordinary convolution operations to extract features, and the last three stages introduce large kernel convolutions to enhance the recognition of complex targets; the feature-level super-resolution branch module is used to reintroduce the original image details of small targets through medium and high-level feature super-resolution, thereby improving the detection of small-sized targets; the multi-scale feature fusion module uses a feature pyramid structure to output feature maps of different scales; the detection module detects the generated feature maps and generates predicted anchor boxes and categories.
[0042] The detection process of the method is as follows: Figure 1 As shown, the following steps are included:
[0043] Step 1: Construct a defect detection network model, including a sequentially connected feature extraction backbone module, a feature-level super-resolution branch module, a multi-scale feature fusion module, and a detection module.
[0044] The feature extraction backbone module consists of four convolutional feature extraction stages. The first stage uses ordinary convolution operations to extract features and downsamples the image twice by a factor of 2. Then, in the second, third, and fourth stages, a large-kernel convolution feature extraction module is introduced to introduce a larger receptive field to enhance the extraction of complex defect features, and the image features are downsampled by a factor of 2 in sequence.
[0045] Among them, the large-core convolution feature extraction module uses a larger convolution kernel to build long-range dependencies and enhance the extraction and understanding of image context information. Its core operation is the large-core separable convolution as follows:
[0046] X dw =Conv 1×k (Conv k×1 (X)
[0047] X ddw =Conv 1×K (Conv K×1 (X dw ))
[0048] X lsk =LSK(X)=X+X ddw
[0049] Among them, X is the input feature map of the large kernel separable convolution, Conv 1×k and Conv k×1The kernels are 1×k and k×1 respectively, with deep convolution DW-Conv and Conv 1×K and Conv K×1 They are respectively the deep dilated convolution DDW-Conv with kernels of 1×K and K×1, and the dilation factor is 2; lsk Output features of the large core feature extraction module, X dw is the output feature of the deep convolution extraction module, X ddw It is the output feature of the deep dilated convolution extraction module, and LSK is the large kernel separable convolution operation.
[0050] In order to enhance the interaction between channels, two channel interaction modules consisting of ordinary convolution with a kernel of 1×1 and ReLu activation function are added before and after the large-core separable convolution operation. At the same time, in order to further enhance the image feature extraction capability of the large-core convolution feature extraction module, the squeeze and excitation operation SE is introduced after the second channel interaction module. At this point, the structure of the large-core convolution feature extraction module is as follows:
[0051] X out =SE(CI(LSK(CI(X))))
[0052] Where CI is the channel interaction module, SE is the squeeze and excitation operation, LSK is the large kernel separable convolution operation, and X is the input feature map of the large kernel separable convolution.
[0053] In order to improve the recognition accuracy of small-sized targets, the feature-level super-resolution branch module fuses the mid-level features and high-level features through the multi-scale feature fusion module, and then super-resolves the fused features to make the super-resolved image continuously approach the original image, thereby introducing the detail information in the original image into the mid-level and high-level features.
[0054] In the multi-scale feature fusion module, a series of CI modules are used to fuse the output features of the second and fourth stages of the feature extraction backbone module:
[0055] f * =Fusion(f 2 +f 4 )
[0056] Fusion is the feature fusion module, f 2 and f 4 are the output features of the second and fourth stages of the feature extraction backbone module, respectively, and f * The fused features.
[0057] In order not to bring additional computational burden to the detection network, the feature-level super-resolution branch module uses a large-kernel convolution feature extraction module to achieve better results with fewer module repetitions:
[0058] f lb =LB N (…LB 2 (LB 1 (f * )))
[0059] Among them, LB is a large kernel convolution feature extraction module, f lb is the output feature of the Nth large kernel convolution feature extraction module. Finally, in order to obtain the final super-resolution image, f * and f lb are simultaneously input into the reconstruction module:
[0060] SR=Rec(f lb )+Rec(f * )
[0061] Among them, Rec is the reconstruction module, SR is the super-resolution result, and f * is the fused feature, f lb It is the output feature of the Nth large kernel convolution feature extraction module.
[0062] The first stage of the feature extraction backbone module contains multiple convolutional feature extraction modules to fully extract detailed texture information. The second, third and fourth stages introduce large-core convolutional feature extraction modules to obtain a larger receptive field, thereby enhancing the extraction of global context information. Each stage of the feature extraction backbone module outputs a feature map and sends it to the multi-scale feature fusion module for information aggregation.
[0063] Step 2: Data preprocessing, obtain RGB images of various defect types in industrial scenes. Before inputting the defect detection network model training, each RGB image is first scaled to 1024 pixels on the short side while maintaining the aspect ratio, and then randomly scaled left and right and up and down by 50%, and the manual marking results are also reversed accordingly. Next, the RGB image pixels are normalized from 0-255 to the 0-1 range, and finally all RGB images are padded with peripheral pixels so that the RGB image size is 1024*1024 pixels, and the data set is obtained. The data set is randomly divided into a training set and a test set. The collected RGB images are transmitted to the computer that executes the algorithm.
[0064] Step 3: Training process. After the image is input into the defect detection network model, it first passes through the feature extraction backbone module to obtain 4 layers of effective feature maps. The feature-level super-resolution branch module synchronously trains the 4 layers of effective feature maps to enhance the detail information of small targets of higher-level features. Then, the multi-scale feature fusion module performs information fusion and outputs the 3-layer fused feature map. Finally, the detection module outputs the defect location and classification results, and the loss function is calculated based on this result and the result of manual calibration.
[0065] The total loss function consists of two parts:
[0066] L total =L de +αL sr ;
[0067] In the formula, α is the super-resolution loss L sr The weight coefficient, L de To detect losses;
[0068] Among them, the super-resolution loss L sr By calculating the loss L1 between the input image I and the super-resolution result SR, we get:
[0069] L sr =||SR-I|| 1
[0070] In the formula, the input image I is the original RGB image to be super-resolution processed, and the super-resolution result SR is the high-resolution image obtained by processing the low-resolution image with the super-resolution algorithm;
[0071] The super-resolution loss is calculated by comparing the difference between the high-resolution image obtained after the input low-resolution image is processed by the super-resolution algorithm and the ideal high-resolution image.
[0072] In each training step, back propagation is performed starting from the loss function value, and the Adam optimizer is used to optimize the network parameters according to the gradient information obtained by back propagation, thereby guiding the neural network to achieve accurate defect detection results based on the input image.
[0073] The large core feature extraction module described in step 1 performs feature extraction in a serial manner; the feature-level super-resolution branch module uses the super-resolution branch to fuse the mid- and high-level features, and uses the proposed super-resolution method to super-resolve the fused features, reintroducing the original RGB image detail information into the mid- and high-level information.
[0074] Example 1
[0075] The industrial engine surface defect dataset ESD used in this experiment contains a total of 497 RGB color images of the surface of engine parts taken in real industrial scenes, with an image size of 3700*3620 pixels. It is randomly shuffled and divided into a training set consisting of 448 images and a test set consisting of 49 images. The images will first be scaled to 1024*1024 pixels in the preprocessing stage. The initialization method of the defect detection network parameters in the present invention is that the parameters of the feature extraction module are all randomly initialized. The operating environment is a computer with a PyTorch framework that can read a given image and complete the construction and training of the model of this method. The training time of this embodiment is about 3 hours on a Gold 6626R@2.90GHz CPU, 8G memory and NVIDIAGeForce RTX3090 GPU.
[0076] First, relevant training parameters are set, and the optimizer used for network update in the present invention is set to Adam optimizer, its momentum value is set to 0.9, the initial learning rate is set to 0.001, and the weight decay coefficient is set to 0.0001.
[0077] Then, the defect detection network model proposed in the present invention is constructed. Figure 2 The structure of the algorithm is shown in Figure 2, which consists of a feature extraction backbone module with 4 stages, a feature-level super-resolution branch module, a multi-scale feature fusion module and a detection module. Figure 3 As shown in Figure 2, the large kernel convolution feature extraction module uses deep separable convolution to obtain a larger receptive field with fewer parameters, as shown in Figure 2. Figure 4 As shown in Figure 2, the feature-level super-resolution branch first fuses the features of the second and fourth stages, and then inputs the fused features into the super-resolution module for super-resolution to enhance the small target information in the features. Finally, the enhanced multi-scale features are passed to the detection module to obtain the detection result image with rectangular anchor box markers.
[0078] Next, when using the divided data set for network training, 4 images are randomly selected from the training set each time to input into the network, and the selected stochastic gradient descent optimizer is used to update the parameters. The training is completed after 100 rounds of iterations on the data set. Finally, the images in the test set are input into the trained network for detection, and the results of the embodiment of the invention are obtained, as shown in FIG. Figure 5 shown.
[0079] In another embodiment of the present invention, a computer device is provided, the computer device comprising a processor and a memory, the memory being used to store a computer program, the computer program comprising program instructions, and the processor being used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, which are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding functions; the processor described in the embodiment of the present invention can be an operation of an industrial defect detection method based on large kernel convolution and feature-level super-resolution.
[0080] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0081] It should be noted that the terms "including" and "having" in the specification and claims of the present invention and the above-mentioned drawings and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.
[0082] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any ordinary technician in the industry can smoothly implement the present invention as shown in the drawings and described above. However, any equivalent changes, modifications and evolutions made by technicians familiar with the profession without departing from the scope of the technical solution of the present invention using the technical content disclosed above are all equivalent embodiments of the present invention. At the same time, any equivalent changes, modifications and evolutions made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the technical solution of the present invention.
Claims
1. An industrial defect detection method based on large kernel convolution and feature-level super-resolution, characterized in that: include, Constructing a defect detection network model, the defect detection network model includes a feature extraction backbone module, a feature-level super-resolution branch module, a multi-scale feature fusion module and a detection module; a large-kernel convolution feature extraction module is introduced into the feature extraction backbone module; Obtain images of various defect types in industrial scenes, perform data preprocessing on the images, and obtain training sets and test sets; The defect detection network model is trained based on the training set, and the test set is input into the trained defect detection network model for testing to obtain the final defect detection network model; The industrial image to be inspected is input into the final defect detection network model to detect defects in the industrial image.
2. The industrial defect detection method based on large kernel convolution and feature-level super-resolution according to claim 1, characterized in that: The feature extraction backbone module includes four convolution feature extraction stages, wherein the first stage uses conventional convolution operations to extract features and downsamples the image twice by a factor of 2; the second, third, and fourth stages introduce large-kernel convolution feature extraction modules to enhance the extraction of complex defect features, and downsample the image features once by a factor of 2, respectively.
3. The industrial defect detection method based on large kernel convolution and feature-level super-resolution according to claim 2, characterized in that: Two channel interaction modules consisting of conventional convolution with a kernel of 1*1 and a ReLu activation function are added to the large-kernel convolution feature extraction module.
4. The industrial defect detection method based on large kernel convolution and feature level super resolution according to claim 1, characterized in that: The feature-level super-resolution branch module fuses the middle-level features and the high-level features through the multi-scale feature fusion module, and then super-resolves the fused features so that the super-resolved image continuously approaches the original image, thereby introducing the detail information in the original image into the middle and high-level features.
5. The industrial defect detection method based on large kernel convolution and feature-level super-resolution according to claim 1, characterized in that: The multi-scale feature fusion module uses a feature pyramid structure to output feature maps of different scales.
6. The industrial defect detection method based on large kernel convolution and feature level super resolution according to claim 1, characterized in that: The image is preprocessed, the specific process is as follows: Each image is scaled to 1024 pixels on the short side while maintaining its aspect ratio. Then, it is randomly scaled left and right and up and down by 50% and the manually labeled results are inverted accordingly. Then, the image pixels are normalized from 0-255 to 0-1. Finally, all images are padded with peripheral pixels to make the image size 1024*1024 pixels. The image dataset is obtained and then randomly divided into training and test sets.
7. The industrial defect detection method based on large kernel convolution and feature level super resolution according to claim 1, characterized in that: The defect detection network model is trained based on the training set. The specific process is as follows: The images in the training set are input into the defect detection network model, and are first processed by the feature extraction backbone module to obtain 4 layers of effective feature maps. The 4 layers of effective feature maps are synchronously trained through the feature-level super-resolution branch module to enhance the detail information of small targets of higher-level features. Then, the multi-scale feature fusion module is used to fuse the information and output the 3-layer fused feature maps. Finally, the detection module outputs the defect location and classification results. The total loss function is calculated based on the defect location and classification results and the results of manual calibration. Back propagation is performed through the total loss function, and the Adam optimizer is used to optimize the parameters in the defect detection network model according to the gradient information obtained by back propagation, thereby obtaining the trained defect detection network model.
8. The industrial defect detection method based on large kernel convolution and feature level super resolution according to claim 7, characterized in that: The specific calculation formula of the total loss function is: In the formula, L total is the total loss function, L sr is the excess loss, Excess loss The weight coefficient of To detect losses; Among them, the excess loss By calculating the input image And super-resolution results The loss value L1 between them is obtained, and the specific formula is: 。 9. An industrial defect detection system based on large kernel convolution and feature-level super-resolution, characterized in that: The industrial defect detection method based on large kernel convolution and feature-level super-resolution according to any one of claims 1 to 8 comprises: The data preprocessing module is used to obtain images of various defect types in industrial scenes, and obtain training sets and test sets after data preprocessing of the images; A model building module, used to build a defect detection network model, wherein the defect detection network model includes a feature extraction backbone module, a feature-level super-resolution branch module, a multi-scale feature fusion module and a detection module; The model training module is used to train the defect detection network model with the training set, input the test set into the trained defect detection network model, perform the test, and obtain the final defect detection network model; The industrial image detection module is used to input the industrial image to be detected into the final defect detection network model to detect defects in the industrial image.
10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of an industrial defect detection method based on large kernel convolution and feature-level super-resolution as described in any one of claims 1 to 8 are implemented.