Product surface defect identification method and device based on multi-domain feature fusion network, and medium

Abstract

The invention relates to a product surface defect identification method and device based on a multi-domain feature fusion network, and a medium, and the method comprises the steps: inputting a defect image obtained by collecting a product surface defect identification data set into a first-layer network of an A-ACN backbone network, and obtaining a first feature map; inputting the defect image into a peak region auxiliary feature learning module to obtain a peak feature map, and adding the peak feature map and the first feature map to obtain a second feature map; inputting the second feature map into a second-layer network of the backbone network to obtain a third feature map, and respectively inputting the third feature map and the defect image into a multi-domain mixed attention module to obtain a fourth feature map; inputting the fourth feature map into a third-layer network of the backbone network, and obtaining a fifth feature map through a gating feature selection module; and inputting the fifth feature map into a fourth-layer network and a fifth-layer network of the backbone network to obtain a final output defect category. Compared with the prior art, the method has the advantages of high precision, strong generalization ability, strong robustness and the like.

Description

Technical Field

[0001] The present invention relates to the field of product surface defect recognition, and in particular to a product surface defect recognition method, equipment and medium based on a multi-domain feature fusion network. Background Art

[0002] As the manufacturing industry shifts from quantity expansion to quality improvement, improving product quality has become the key to improving product competitiveness. Since most products are mass-produced, the production process involves many complex process flows, and any abnormality in the manufacturing process may lead to defects on the product surface. However, surface damage to some key components will not only cause serious damage to the performance of the product, but may also pose safety risks to users. Surface defect recognition methods can detect problems in a timely manner, eliminate unqualified products in a timely manner, improve the corresponding production process, and maintain production machines predictively. Through automated and intelligent surface defect detection methods, the causes of defects are analyzed, making the maintenance work of the production line more targeted, effectively reducing the product failure rate, improving product quality, and ensuring user safety.

[0003] Traditional product surface defect recognition methods are mostly implemented through traditional manual feature extraction and classifier design, usually using interpretable mathematical theories and mature expert experience to analyze the representative features of product surface defects. Usually, such algorithms rely on rich expert knowledge and theoretical models, and it is difficult to adapt to complex and changing scenarios and the needs of intelligent recognition. In recent years, research on product surface defect recognition based on deep learning methods has gradually become mainstream in this field, and convolutional neural network models have been applied to the field of product surface defect recognition. Although this data-driven method can automatically learn hidden patterns and characteristics from large-scale sample data, it has higher accuracy, faster speed and simpler design ideas than traditional methods. However, in the actual application of product surface defect recognition, it usually faces problems such as difficult to explain, small samples of big data, and limitations of visual cognition. Deep integration of theory-driven and data-driven methods is particularly important in the field of product surface defect recognition. A method that improves the accuracy of product surface defect recognition and model interpretability by integrating traditional features and deep features has been proposed. However, there are still some problems to be solved in the actual product surface defect scene. First, product surface defect images usually have complex and changeable shapes and texture structures, so it is necessary to consider how to better capture the features of key areas in limited sample learning. Secondly, many studies integrate physical models into deep neural networks for feature fusion, but because the model performance is greatly affected by fixed input physical features, when faced with new samples that are significantly different from the distribution of training data, its generalization ability may be limited and its adaptability is not strong. In addition, such algorithms usually have a complex construction process and high demand for computing resources. In addition, fine-grained model methods can directly improve the recognition performance of deep learning models, but it is difficult to achieve a good balance between model accuracy and speed in actual application scenarios. Therefore, how to improve defect recognition capabilities, enhance adaptability and generalization capabilities, and improve classification accuracy in product surface defect recognition is a technical problem that needs to be solved. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a product surface defect recognition method, equipment and medium based on a multi-domain feature fusion network. By integrating the peak area auxiliary feature learning module, the multi-domain hybrid attention module and the gated feature selection module with the A-CAN backbone network, the defect recognition capability is improved and the characteristics of the image target are comprehensively described. The output of the multi-domain feature fusion is used for product surface defect recognition.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] According to one aspect of the present invention, a method for identifying product surface defects based on a multi-domain feature fusion network is provided, and the specific steps include:

[0007] S1. Collect product surface defect recognition data sets and perform preprocessing to obtain defect images with uniform pixel size;

[0008] S2, inputting the defect image into the first layer network of the A-ACN backbone network to obtain a first feature map;

[0009] S3, inputting the defect image into a peak region auxiliary feature learning module to obtain a peak feature map, and adding the peak feature map to the first feature map to obtain a second feature map;

[0010] S4, input the second feature map into the second layer network of the A-ACN backbone network to obtain a third feature map, input the third feature map and the defect image into the multi-domain hybrid attention module respectively, and obtain a fourth feature map through the wavelet drive branch and the attention branch;

[0011] S5, inputting the fourth feature map into the third layer network of the A-ACN backbone network, and obtaining the fifth feature map through the gated feature selection module;

[0012] S6. Input the fifth feature map into the fourth and fifth network layers of the A-ACN backbone network to obtain the final output defect category.

[0013] Furthermore, in S3, the step of using the peak area auxiliary feature learning module to obtain the peak feature map includes: performing Gaussian filtering on the defect image to retain the main features; adaptively selecting the optimal threshold through the OTSU method; obtaining the maximum connected area, that is, the target local peak area, through the binary threshold method; generating a binary map of the target local peak area as the peak feature map.

[0014] Furthermore, the expression of the Gaussian filter is:

[0015]

[0016] Wherein, (x, y) is the point coordinate, σ is the standard deviation, k is the window template size parameter, and the window template size is (2k+1)×(2k+1).

[0017] Furthermore, the binary threshold method is expressed as:

[0018]

[0019] Among them, dst(x, y) is the pixel value of the target image at the coordinate position (x, y), src(x, y) is the pixel of the source image at the coordinate position (x, y), thresh is the threshold automatically determined by the Otsu method, and maxval is the maximum pixel value.

[0020] Furthermore, in S4, the feature map value is re-weighted based on the global visual information through the frequency domain attention module; the low-frequency information A, horizontal high-frequency information H, vertical high-frequency information V and diagonal high-frequency information D of the defect image are obtained through the two-dimensional discrete wavelet transform of the wavelet-driven branch, and the eigenvalue expression of the (i, j)th wavelet feature map is:

[0021] x A (i,j)=x(2i-1,2j-1)+x(2i-1,2j)+x(2i,2j-1)+x(2i,2j),

[0022] x H (i,j)=-x(2i-1,2j-1)-x(2i-1,2j)+x(2i,2j-1)+x(2i,2j),

[0023] x V (i,j)=-x(2i-1,2j-1)+x(2i-1,2j)-x(2i,2j-1)+x(2i,2j),

[0024] x D (i,j)=x(2i-1,2j-1)-x(2i-1,2j)-x(2i,2j-1)+x(2i,2j),

[0025] Among them, x A (i, j) is the low-frequency information of the (i, j)th wavelet feature map, x H (i, j) is the horizontal high-frequency information of the (i, j)th wavelet feature map, x V (i, j) is the vertical high-frequency information of the (i, j)th wavelet feature map, x D (i, j) is the diagonal high-frequency information of the (i, j)th wavelet feature map.

[0026] Furthermore, the features of the obtained wavelet feature map are processed by group convolution to retain the local semantics of each wavelet feature and generate a wavelet activation feature map; the third feature map is convolved to make it consistent with the dimension of the wavelet activation feature map, and the element-by-element product of the third feature map and the wavelet activation feature map is obtained as the wavelet activation group feature that emphasizes each wavelet component; the compressed features of each subgroup are obtained by the selective average pooling layer, and the output is obtained by processing with the MLP layer; after the subgroup features are spliced, the channel weights are recalibrated by the Softmax layer; and they are multiplied channel by channel with the wavelet activation group features to obtain the local semantic enhancement features of the wavelet-driven branch based on wavelet information.

[0027] Furthermore, in S5, the fourth feature map is input into a gated feature selection module, and the channel is divided into a gated path and a convolutional path, wherein the gated path obtains a weight ranging from 0 to 1 through a sigmoid function, and the higher the weight, the more important the feature is at that position; the convolutional path nonlinearly activates the feature map through a LeakyReLU layer; and the output results of the gated path and the convolutional path are multiplied to obtain a fifth feature map.

[0028] Furthermore, during the execution of the method, the training parameters of the model are updated through back propagation of the cross entropy loss function.

[0029] According to a second aspect of the present invention, there is provided an electronic device, comprising a memory and a processor, wherein a computer program is stored in the memory, and the method described above is implemented when the processor executes the program.

[0030] According to a third aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, wherein the program implements the method described above when executed by a processor.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] (1) Improving the accuracy of product surface defect recognition: By introducing the peak area auxiliary feature learning module, the network's learning of target key areas, especially target shape and structure features, is effectively strengthened, and the model's ability to capture product surface defect images with complex and changeable shapes and texture structures is improved, so that key features can be accurately identified even in limited sample learning. In addition, the wavelet-driven multi-domain hybrid attention module adaptively fuses semantic prior knowledge, wavelet domain and frequency domain attention features, and then uses the gated feature selection mechanism to filter and enhance the features, activate and reweight the features, and strengthen the learning of local semantics, so as to have a deeper understanding of the defect information in the image, enhance the model's generalized feature extraction capability, and improve the accuracy of defect recognition.

[0033] (2) Improving the robustness of recognition: Through the gated feature selection module, the features are effectively filtered and enhanced. The gated feature selection mechanism can focus on relevant features and suppress irrelevant features, thereby enhancing the significance of the learned features. This helps the model focus more on key defect features during the recognition process, thereby improving the accuracy and robustness of recognition.

[0034] (3) Enhanced generalization capability: The peak area auxiliary feature learning module, multi-domain hybrid attention module and gated feature selection module are fully integrated with the A-ACN backbone network to construct a multi-domain feature fusion network, which fully extracts the spatial domain, wavelet domain and frequency domain features of the image, and optimizes the fusion of high-order deep features and low-order local artificial features to comprehensively describe the characteristics of the image target, so that the model shows higher performance, stronger generalization capability and credibility in product surface defect recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flowchart of the steps of the product surface defect recognition method based on the multi-domain feature fusion network;

[0036] Figure 2 A schematic diagram of the network structure for product surface defect recognition based on a multi-domain feature fusion network; DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are 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.

[0038] like Figure 1 As shown in FIG. 1 , a product surface defect recognition method based on a multi-domain feature fusion network is shown, and the specific steps include:

[0039] S1. Collect product surface defect recognition data sets and perform preprocessing to obtain defect images with uniform pixel size;

[0040] S2, input the defect image into the first layer network of the A-ACN backbone network to obtain the first feature map;

[0041] S3, inputting the defect image into the peak region auxiliary feature learning module to obtain a peak feature map, and adding the peak feature map to the first feature map to obtain a second feature map;

[0042] S4, input the second feature map into the second layer network of the A-ACN backbone network to obtain a third feature map, input the third feature map and the defect image into the multi-domain hybrid attention module respectively, and obtain a fourth feature map through the wavelet drive branch and the attention branch;

[0043] S5, inputting the fourth feature map into the third layer network of the A-ACN backbone network, and obtaining the fifth feature map through the gated feature selection module;

[0044] S6. Input the fifth feature map into the fourth and fifth network layers of the A-ACN backbone network to obtain the final output defect category.

[0045] This embodiment uses the WM-811K wafer surface image dataset collected during the actual semiconductor production process for experimental verification. The WM-811K dataset contains a total of 811,457 wafer surface images with marked defect patterns, including 9 wafer patterns, covering 8 defect patterns including center defects, circular rings, edge local defects, edge ring defects, local defects, near-full defects, random defects, scratches, and normal wafer surface patterns. 32,077 wafer images were randomly selected from the WM-811K dataset to form the data used in this article. The images were preprocessed to a size of 128×128 pixels, and each type of data was divided into a training set and a test set as network input in a ratio of 7:3. The training set is used to train MDFFN, and the network parameters are adjusted. Finally, the test set is used to evaluate the classification effect of MDFFN. The backbone network is used to extract features from the input product surface defect image. A-ConvNets (ACN) is a classic deep learning algorithm in the field of target recognition. The backbone network is a fully convolutional structure. The backbone network used in this embodiment adjusts the parameters and structure based on the original ACN algorithm, which is denoted as A-ACN. The input is a product surface defect image. The backbone network extracts product surface defect image features through five stages {F1, F2, F3, F4, F5}. Each stage i contains a convolution layer for feature extraction and a maximum pooling layer for feature map dimensionality reduction. The output feature map is represented as {E1, E2, E3, E4, E5}. In addition, each convolution layer output is batch normalized and activated with a rectified linear unit to increase network nonlinearity. The spatial size of the feature map in the latter stage is reduced to half of that in the previous stage. As the network depth increases, the convolution kernel channel width is expanded by a multiple of 2, so that features with stronger abstract semantics can be extracted.

[0046] In S3, a peak region auxiliary feature learning mechanism is constructed to extract the peak features of the image and fuse them with the deep features. After the target peak map is embedded in the first layer of the network structure, the network prompts neurons to respond actively to the peak region during the learning optimization process, which can strengthen the learning of the key target areas, especially the target shape and structure features. The steps of using the peak region auxiliary feature learning module to obtain the peak feature map include: Gaussian filtering the defect image to retain the main features; adaptively selecting the optimal threshold through the OTSU method; obtaining the maximum connected area, that is, the target local peak area through the binary threshold method; generating a binary map of the target local peak area as the peak feature map.

[0047] The expression of Gaussian filtering is:

[0048]

[0049] Wherein, (x, y) is the point coordinate, σ is the standard deviation, k is the window template size parameter, and the window template size is (2k+1)×(2k+1).

[0050] The expression of binary threshold method is:

[0051]

[0052] Among them, dst(x, y) is the pixel value of the target image at the coordinate position (x, y), src(x, y) is the pixel of the source image at the coordinate position (x, y), thresh is the threshold automatically determined by the Otsu method, and maxval is the maximum pixel value.

[0053] After the first layer of the network outputs the first feature map and the peak feature map, each undergoes a 3×3 convolution process and is converted into the same dimension. The local peak feature map is multiplied by a stable and slowly decreasing adjustment factor α(t) and then added to the network feature map before being passed to the next layer. The expression is:

[0054]

[0055] Among them, Q is the peak map after convolution processing, E1 is the first feature map after convolution processing, and the two feature maps are added to obtain the output feature map α(t) is a regulation factor based on the change of the number of training rounds t, which gradually decreases during the training process and is expressed as:

[0056]

[0057] Among them, t is the current training round number, t max is the maximum number of training rounds.

[0058] In S4, the feature map values ​​are re-weighted based on the global visual information through the frequency domain attention module; the low-frequency information A, horizontal high-frequency information H, vertical high-frequency information V and diagonal high-frequency information D of the defect image are obtained through the two-dimensional discrete wavelet transform of the wavelet-driven branch. The eigenvalue expression of the (i, j)th wavelet feature map is:

[0059] x A (i,j)=x(2i-1,2j-1)+x(2i-1,2j)+x(2i,2j-1)+x(2i,2j),

[0060] x H (i,j)=-x(2i-1,2j-1)-x(2i-1,2j)+x(2i,2j-1)+x(2i,2j),

[0061] x V (i,j)=-x(2i-1,2j-1)+x(2i-1,2j)-x(2i,2j-1)+x(2i,2j),

[0062] x D (i,j)=x(2i-1,2j-1)-x(2i-1,2j)-x(2i,2j-1)+x(2i,2j),

[0063] Among them, x A (i, j) is the low-frequency information of the (i, j)th wavelet feature map, x H (i, j) is the horizontal high-frequency information of the (i, j)th wavelet feature map, x V (i, j) is the vertical high-frequency information of the (i, j)th wavelet feature map, x D (i, j) is the diagonal high-frequency information of the (i, j)th wavelet feature map.

[0064] The features of the obtained wavelet feature map are processed by group convolution to retain the local semantics of each wavelet feature and generate a wavelet activation feature map; the third feature map is convolved to make it consistent with the dimension of the wavelet activation feature map, and the element-by-element product of the third feature map and the wavelet activation feature map is obtained as the wavelet activation group feature that emphasizes each wavelet component; the compressed features of each subgroup are obtained by the selective average pooling layer, and the output is obtained by processing with the MLP layer; after the subgroup features are spliced, the channel weights are recalibrated by the Softmax layer; and they are multiplied channel by channel with the wavelet activation group features to obtain the local semantic enhancement features of the wavelet-driven branch based on wavelet information.

[0065] Generate a wavelet activation feature map based on the feature pass of the wavelet feature map The expression is:

[0066]

[0067] Among them, Conv i is the i-th group convolutional layer, W i is the i-th wavelet feature;

[0068] The element-by-element product expression of the third feature map and the wavelet activation feature map is:

[0069]

[0070] in, The third feature map is unified with the wavelet activation feature map into a feature map of the same dimension after convolution processing.

[0071] In S5, the fourth feature map is input into the gated feature selection module, and the channel is divided into a gated path and a convolution path, wherein the gated path obtains a weight ranging from 0 to 1 through a sigmoid function, and the higher the weight, the more important the feature is at this position; the convolution path nonlinearly activates the feature map through a LeakyReLU layer; the output results of the gated path and the convolution path are multiplied to obtain the fifth feature map O y,x , the expression is:

[0072]

[0073] Gating y,x =ΣΣC g E3,

[0074] Conv y,x =ΣΣC f E3,

[0075] Among them, E4 is the fourth characteristic map, C g and C f is the convolution operation, σ is the sigmoid function, which transforms the gated value output to between 0 and 1. is the LeakyReLU activation function.

[0076] During the execution of the method, the training parameters of the model are updated through back propagation of the cross entropy loss function.

[0077] like Figure 2 As shown in the figure, it is a schematic diagram of the network structure of product surface defect recognition based on multi-domain feature fusion network. The peak area auxiliary feature learning module, multi-domain hybrid attention module and gated feature selection module are integrated with A-ACN to fully extract the spatial domain, wavelet domain and frequency domain features of the image to optimize the fusion of high-order deep features and low-order local artificial features, thereby comprehensively describing the characteristics of the image target.

[0078] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0079] The electronic device of the present invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0080] A plurality of components in the device are connected to the I / O interface, including: an input unit, such as a keyboard, a mouse, etc.; an output unit, such as various types of displays, speakers, etc.; a storage unit, such as a disk, an optical disk, etc.; and a communication unit, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit allows the device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks. The processing unit performs the various methods and processes described above, such as the method of the present invention. For example, in some embodiments, the method of the present invention may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via a ROM and / or a communication unit. When the computer program is loaded into the RAM and executed by the CPU, one or more steps of the method of the present invention described above may be performed. Alternatively, in other embodiments, the CPU may be configured to perform the method of the present invention in any other appropriate manner (e.g., by means of firmware).

[0081] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0082] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0083] In the context of the present invention, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0084] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A product surface defect recognition method based on a multi-domain feature fusion network, characterized in that: The specific steps include: S1. Collect product surface defect recognition data sets and perform preprocessing to obtain defect images with uniform pixel size; S2, inputting the defect image into the first layer network of the A-ACN backbone network to obtain a first feature map; S3, inputting the defect image into a peak region auxiliary feature learning module to obtain a peak feature map, and adding the peak feature map to the first feature map to obtain a second feature map; S4, input the second feature map into the second layer network of the A-ACN backbone network to obtain a third feature map, input the third feature map and the defect image into the multi-domain hybrid attention module respectively, and obtain a fourth feature map through the wavelet drive branch and the attention branch; S5, inputting the fourth feature map into the third layer network of the A-ACN backbone network, and obtaining the fifth feature map through the gated feature selection module; S6. Input the fifth feature map into the fourth and fifth network layers of the A-ACN backbone network to obtain the final output defect category.

2. The method for identifying product surface defects based on a multi-domain feature fusion network according to claim 1, characterized in that: In S3, the step of using the peak area auxiliary feature learning module to obtain the peak feature map includes: performing Gaussian filtering on the defect image to retain the main features; adaptively selecting the optimal threshold through the OTSU method; obtaining the maximum connected area, that is, the target local peak area, through the binary threshold method; generating a binary map of the target local peak area as the peak feature map.

3. The method for identifying product surface defects based on a multi-domain feature fusion network according to claim 2 is characterized in that: The expression of the Gaussian filter is: Wherein, (x, y) is the point coordinate, σ is the standard deviation, k is the window template size parameter, and the window template size is (2k+1)×(2k+1).

4. The method for identifying product surface defects based on a multi-domain feature fusion network according to claim 2 is characterized in that: The expression of the binary threshold method is: Among them, dst(x, y) is the pixel value of the target image at the coordinate position (x, y), src(x, y) is the pixel of the source image at the coordinate position (x, y), thresh is the threshold automatically determined by the Otsu method, and maxval is the maximum pixel value.

5. The method for identifying product surface defects based on a multi-domain feature fusion network according to claim 1, characterized in that: In S4, the feature map value is re-weighted based on the global visual information through the frequency domain attention module; the low-frequency information A, horizontal high-frequency information H, vertical high-frequency information V and diagonal high-frequency information D of the defect image are obtained through the two-dimensional discrete wavelet transform of the wavelet-driven branch, and the eigenvalue expression of the (i, j)th wavelet feature map is: <h2 style=";text-align:left;direction:ltr">x<h2 style=";text-align:left;direction:ltr"> A <h2 style=";text-align:left;direction:ltr"> (i,j) = x(2i-1,2j-1)+x(2i-1,2j)+x(2i,2j-1)+x(2i,2j), <h2 style=";text-align:left;direction:ltr">x<h2 style=";text-align:left;direction:ltr"> H <h2 style=";text-align:left;direction:ltr"> (i,j) = -x(2i - 1, 2j - 1) - x(2i - 1, 2j) + x(2i, 2j - 1) + x(2i, 2j) <h2 style=";text-align:left;direction:ltr">x<h2 style=";text-align:left;direction:ltr"> V <h2 style=";text-align:left;direction:ltr"> (i,j) = -x(2i - 1, 2j - 1) + x(2i - 1, 2j) - x(2i, 2j - 1) + x(2i, 2j) <h2 style=";text-align:left;direction:ltr">x<h2 style=";text-align:left;direction:ltr"> D <h2 style=";text-align:left;direction:ltr"> (i,j) = x(2i-1,2j-1)-x(2i-1,2j)-x(2i,2j-1)+x(2i,2j), Among them, x A (i, j) is the low-frequency information of the (i, j)th wavelet feature map, x H (i, j) is the horizontal high-frequency information of the (i, j)th wavelet feature map, x V (i, j) is the vertical high-frequency information of the (i, j)th wavelet feature map, x D (i, j) is the diagonal high-frequency information of the (i, j)th wavelet feature map.

6. The method for identifying product surface defects based on a multi-domain feature fusion network according to claim 5, characterized in that: The features of the obtained wavelet feature map are processed by group convolution to retain the local semantics of each wavelet feature and generate a wavelet activation feature map; the third feature map is processed by convolution to make it consistent with the dimension of the wavelet activation feature map, and the element-by-element product of the third feature map and the wavelet activation feature map is obtained as the wavelet activation group feature that emphasizes each wavelet component; The compressed features of each subgroup are obtained through the selective average pooling layer, and the output is obtained by processing through the MLP layer. After the subgroup features are concatenated, the channel weights are recalibrated by the Softmax layer. The features are multiplied channel by channel with the wavelet activation group features to obtain the local semantic enhancement features of the wavelet-driven branch based on wavelet information.

7. The method for identifying product surface defects based on a multi-domain feature fusion network according to claim 1, characterized in that: In S5, the fourth feature map is input into the gated feature selection module, and the channel is divided into a gated path and a convolution path, wherein the gated path obtains a weight ranging from 0 to 1 through a sigmoid function, and the higher the weight, the more important the feature is at this position; the convolution path nonlinearly activates the feature map through a LeakyReLU layer; and the output results of the gated path and the convolution path are multiplied to obtain the fifth feature map.

8. The method for identifying product surface defects based on a multi-domain feature fusion network according to claim 1, characterized in that: During the execution of the method, the training parameters of the model are updated through back propagation of the cross entropy loss function.

9. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

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

Images