A photovoltaic surface defect detection system and method based on image recognition processing

Abstract

The application discloses a photovoltaic surface defect detection system and method based on image recognition processing. The system comprises a feature extraction network, a hybrid encoder and a decoder. The feature extraction network comprises a feature enhancement module adopting a convolution and additive self-attention hybrid structure, which is used for extracting multi-scale features. The hybrid encoder is composed of an intra-scale feature interaction module and a cross-scale feature fusion module. The intra-scale feature interaction module performs global modeling and noise suppression on high-level features through a self-attention unit and a frequency domain discriminative feedforward unit. The cross-scale feature fusion module adopts a fusion strategy based on a gating convolution unit to realize dynamic screening and fusion of multi-scale features. The method corresponds to the operation process of the system. The application is aimed at the characteristics of low contrast, multiple noises and defect superposition of EL images. Through the cooperative optimization of the above modules, the weak defect capture ability, the anti-interference ability and the multi-defect distinguishing ability are improved, and high-precision and high-real-time defect detection is realized.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and artificial intelligence, specifically to a photovoltaic surface defect detection system and method based on image recognition processing. Background Technology

[0002] As the photovoltaic industry expands to a large scale, product quality control faces severe challenges. Electroluminescence imaging (EL) technology, as a key non-destructive testing method, can reveal internal defects in solar cells. However, EL images of photovoltaic cells generally suffer from low basic quality, weak feature contrast, susceptibility to imaging noise, and confusion caused by the superposition of multiple types of defects. Traditional detection methods based on visual inspection or general deep learning models struggle to achieve high accuracy, robustness, and real-time performance simultaneously in such complex scenarios. Existing technologies typically apply general target detection frameworks directly without deep adaptation to the specific physical properties of EL images, resulting in insufficient capture of subtle defects, performance degradation in noisy environments, and poor discrimination of superimposed defects, failing to meet the demands of precise quantitative detection in industrial applications.

[0003] Therefore, this invention proposes a photovoltaic surface defect detection system and method based on image recognition processing. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the prior art, and to propose a photovoltaic surface defect detection system and method based on image recognition processing.

[0005] This invention first discloses a photovoltaic surface defect detection system based on image recognition processing, comprising: The feature extraction network includes multiple sequentially connected feature enhancement modules for extracting features from the input electroluminescent image and outputting a multi-scale feature map. The feature enhancement modules adopt a hybrid structure that integrates convolutional operations and additive self-attention mechanisms. A hybrid encoder, connected to the feature extraction network, is used to receive and process the multi-scale feature map, comprising: The intra-scale feature interaction module is used to perform global context modeling and noise suppression on the received high-level features; A cross-scale feature fusion module, connected to the intra-scale feature interaction module, is used to dynamically filter and fuse features from different scales through a gating mechanism and output a fused feature sequence. A decoder, connected to the hybrid encoder, is used to generate defect detection results based on the fused feature sequence.

[0006] In the above system, the feature enhancement module includes: an integration subnetwork, composed of deep convolutional layers, used to extract local context and perform preliminary enhancement on the input features; A convolutional additive self-attention module is used to perform attention interaction based on additive similarity calculation on the output features of the integrated subnetwork, wherein the additive similarity calculation simultaneously fuses attention information from spatial and channel dimensions; The feedforward subnetwork is used to perform a nonlinear transformation on the output of the convolutional additive self-attention module and to perform a residual connection with the output of the integration subnetwork.

[0007] In the above system, the intra-scale interaction submodule includes: Self-attention units are used to establish global spatial dependencies on the high-level semantic features; The frequency domain filtering feedforward unit is used to convert features to the frequency domain, perform dynamic frequency domain filtering using learnable parameters to suppress noise, and then convert them back to the spatial domain.

[0008] In the above system, the cross-scale feature fusion module adopts a feature fusion strategy based on gated convolutional units. It generates an independent gate signal for each input feature to perform element-level modulation on the corresponding feature values ​​and then completes the fusion.

[0009] Secondly, this invention discloses a photovoltaic surface defect detection method based on image recognition processing, applied to the aforementioned system, the method comprising: The feature extraction network extracts multi-scale feature maps of the input electroluminescent image; The intra-scale feature interaction module in the hybrid encoder performs global context modeling and noise suppression on high-level features in the multi-scale feature map. The cross-scale feature fusion module in the hybrid encoder dynamically fuses noise-suppressed high-level features with features of other scales based on a gating mechanism to obtain a fused feature sequence. The decoder outputs the category and location information of the defects based on the fused feature sequence.

[0010] In the above method, the step of extracting multi-scale feature maps through a feature extraction network includes the following processing steps of the feature enhancement module: Perform depthwise convolution on the input features to integrate local contextual information; For the integrated features, an attention mechanism based on the additive similarity function is used to perform global interactive computation in both spatial and channel dimensions. The features that have undergone attention interaction are subjected to a nonlinear feedforward transformation and then residually connected with the input features before the depthwise convolution operation to output the enhanced features.

[0011] In the above method, the steps of performing global context modeling and noise suppression specifically include: Multi-head self-attention computation is performed on the high-level features to establish global feature associations; Transform the features output by the self-attention calculation to the frequency domain; Dynamically weighted filtering of frequency domain features is performed using a learnable parameter matrix to enhance the target frequency band and suppress noise frequency bands; The filtered frequency domain features are inversely converted back to the spatial domain.

[0012] In the above method, the step of performing dynamic fusion based on the gating mechanism specifically includes: For each scale of input features to be fused, generate the corresponding feature value vector and gate signal vector respectively; The eigenvalue vector is modulated element by element using the gated signal vector; The features at all scales after modulation are aggregated to obtain the fused features.

[0013] Thirdly, the present invention discloses an electronic device, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the methods described above.

[0014] Fourthly, the present invention discloses a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the above-described method. The beneficial effects of this invention are as follows: 1. The feature extraction module employs a feature enhancement unit that fuses convolution and additive self-attention, enhancing the model's ability to capture weak defect features in low-contrast images. This design, by combining detail awareness from local convolution with global dependency modeling through self-attention, can extract richer and more discriminative defect features from blurred electroluminescent images, thereby effectively reducing the false negative rate caused by indistinct features.

[0015] 2. The frequency domain analysis-based noise suppression mechanism introduced in the intra-scale interaction submodule improves the robustness of the model under complex imaging conditions. This module actively identifies and suppresses high-frequency interference introduced by uneven illumination and equipment noise by transforming features to the frequency domain and performing learnable dynamic filtering, making feature representation more stable and reducing false detections and performance fluctuations caused by noise interference.

[0016] 3. The dynamic feature selection strategy based on a gating mechanism adopted in the cross-scale fusion submodule improves the model's accuracy in distinguishing multiple defect superimposed scenarios. This strategy uses gating signals to adaptively weight and modulate features from different scales and sources, strengthening key defect features and suppressing irrelevant or interfering information during feature fusion, thereby reducing confusion between multiple defect features and improving the accuracy of classification and localization. Attached Figure Description

[0017] Figure 1 This is a structural diagram of a photovoltaic surface defect detection system based on image recognition processing disclosed in this invention.

[0018] Figure 2 This is a structural diagram of a feature enhancement module in a photovoltaic surface defect detection system based on image recognition processing disclosed in this invention.

[0019] Figure 3 This is a schematic diagram of the intra-scale feature interaction module in a photovoltaic surface defect detection system based on image recognition processing.

[0020] Figure 4 This is a schematic diagram of the cross-scale feature fusion module in a photovoltaic surface defect detection system based on image recognition processing. Detailed Implementation

[0021] To facilitate understanding of this application and to make the aforementioned objectives, features, and advantages of this application more apparent, a detailed description of specific embodiments of this application is provided below in conjunction with the accompanying drawings. Numerous specific details are set forth in the following description to provide a thorough understanding of this application, and preferred embodiments are shown in the accompanying drawings. However, this application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application. This application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified. In the description of this application, "several" means at least one, such as one, two, etc., unless otherwise explicitly specified. It should be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementations. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is only for describing particular implementations and is not intended to limit the scope of this application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] Reference Figures 1-4 A photovoltaic surface defect detection system based on image recognition processing includes: The feature extraction network includes multiple sequentially connected feature enhancement modules for extracting features from the input electroluminescent image and outputting a multi-scale feature map. The feature enhancement modules adopt a hybrid structure that integrates convolutional operations and additive self-attention mechanisms. A hybrid encoder, connected to the feature extraction network, is used to receive and process the multi-scale feature map, comprising: The intra-scale feature interaction module is used to perform global context modeling and noise suppression on the received high-level features; A cross-scale feature fusion module, connected to the intra-scale feature interaction module, is used to dynamically filter and fuse features from different scales through a gating mechanism and output a fused feature sequence. A decoder, connected to the hybrid encoder, is used to generate defect detection results based on the fused feature sequence.

[0023] The photovoltaic cell electroluminescence image defect detection system provided by this invention has the following overall architecture: Figure 1 As shown. During operation, the system first performs preprocessing such as pixel normalization on the input electroluminescence image of the photovoltaic cell, and then inputs the processed image into a feature extraction network. This feature extraction network uses a deep residual network as its basic framework, but its core building block is not a traditional residual block, but a feature enhancement module specifically designed in this invention.

[0024] In one feasible embodiment, after pixel normalization, the EL image of a photovoltaic cell with a resolution of 640×640 is input into the feature extraction network Backbone for feature extraction. A ResNet series network is used as the basic backbone, retaining the original four-stage feature output structure, but replacing the traditional residual blocks in each stage with convolutional additive self-attention modules. Through the synergistic effect of local convolution and global additive attention, the problems of insufficient low-contrast feature capture and weak global semantic association modeling of traditional convolution are solved. This stage ultimately outputs feature maps at three key scales: P3 (160×160×C3), P4 (80×80×C4), and P5 (40×40×C5), corresponding to different downsampling rates. P3 retains shallow detail features, while P5 contains deep semantic features, providing a rich feature foundation for subsequent multi-scale fusion.

[0025] In traditional residual blocks, feature extraction using 3×3 convolutions and residual connections suffers from limited local receptive fields, making it difficult to capture the global semantic associations of weak defects in low-contrast images and leading to confusion in detection locations. Furthermore, relying solely on single feature representations of spatial domain information fails to simultaneously model key features across channel dimensions, resulting in feature redundancy; background noise and invalid texture information occupying feature dimensions affect the focus of subsequent modules on defect features. To address the shortcomings of traditional convolution in capturing low-contrast features and weak modeling of global semantic associations, this invention designs a feature enhancement module, CAS-Block, to improve the accuracy and efficiency of feature extraction. Figure 2As shown, the feature enhancement module consists of an integration subnetwork, a convolutional additive self-attention module, and a feedforward subnetwork. The integration subnetwork comprises three consecutive depthwise separable convolutional layers with a kernel size of 3×3, responsible for extracting and initially enhancing local contextual information from the input features and filtering out some background noise. Each depthwise separable convolutional layer has a kernel size of 3×3 and a stride of 1. Input features: X in ∈R H×W×C Where H and W are the feature map height and width, and C is the number of channels. After processing by the integrated subnet, the output dimension remains unchanged as a preliminary enhanced feature. This process uses the local feature capture capability of deep convolution to initially filter background noise and enhance local discriminative features such as defect edges and textures, laying the foundation for subsequent attention interaction.

[0026] The convolutional additive self-attention module is the core component of the feature enhancement module CAS-Block. It replaces the traditional dot product similarity of self-attention with an additive similarity function, simultaneously modeling spatial and channel attention to achieve global information association and feature selection. CATM then processes the input features... Three independent 1×1 convolutional layers generate three vectors: query Q, key K, and value V, respectively, where: , ,

[0027] N = H × W represents the number of tokens, d represents the embedding dimension, and finally, channel-dimensional compression and feature space transformation are implemented. The similarity function for additive similarity calculation is:

[0028] Where Φ is the context mapping function, which is calculated from spatial attention and channel attention, and the calculation formula is: , DWConv is a 3×3 depthwise convolution that captures the spatial local dependencies of tokens. CA is a simplified channel attention mechanism that compresses the spatial dimension to 1×1 through global average pooling, and then generates channel weights through 1×1 convolution and a sigmoid activation function, thus achieving the selection of key features in the channel dimension. The attention output is obtained by matrix multiplication of the similarity matrix and the value V, followed by a linear transformation to integrate contextual information. .

[0029] Finally, the output of CATM Enhancement of nonlinear feature representation is achieved through a process consisting of two 1×1 convolutions with a GELU activation function inserted in between. The first 1×1 convolution expands the number of channels to 4C, and the second 1×1 convolution restores the number of channels to C, outputting... With integrated subnet output The final output of the CAS-Block is obtained by adding the residuals together. This process is designed to prevent the gradient vanishing problem and preserve the effective information of the original features.

[0030] The multi-scale feature maps are then fed into the hybrid encoder for further processing. The hybrid encoder consists of two functionally defined and sequentially connected sub-modules. The first sub-module is the intra-scale feature interaction module, whose structure is as follows: Figure 3 As shown, this module primarily processes feature maps from the deepest layers of the feature extraction network, which possess high semantic information but low spatial resolution. It first converts the feature map into a feature sequence through a flattening operation, then establishes a global spatial context association within this sequence using a standard multi-head self-attention mechanism. A key improvement lies in replacing the subsequent feedforward network with a frequency-domain discriminative feedforward unit. This unit first segments the feature map output by self-attention into multiple non-overlapping image patches and performs a Fast Fourier Transform (FFT) on each patch to transform the spatial features to the frequency domain. In the frequency domain, a learnable parameter matrix dynamically weights and filters components of different frequencies, effectively suppressing high-frequency interference components related to imaging noise while retaining low-frequency components related to defect features. After frequency-domain filtering, an Inverse Fourier Transform (IFT) is performed on each image patch to restore the features to the spatial domain, and a patch-folding operation restores the original two-dimensional feature map structure. This process achieves active noise filtering at the feature level, significantly improving the purity and robustness of the features. The self-attention mechanism enables intra-scale interaction of P5 features, primarily to establish global associations of high-dimensional semantic features, providing high-quality feature information for subsequent query selection. However, it has significant limitations when processing photovoltaic cell EL images. The fully connected layer is sensitive to noise; noise interference such as changes in illumination can cause shifts in the feature space distribution, affecting the accuracy of global modeling. Furthermore, during feature interaction, the establishment of global associations relies on a single spatial domain, neglecting key features in the frequency domain, leading to insufficient capture of weak global associations. Therefore, a frequency domain discriminative feedforward network (EDFFN) is introduced to replace the original FFN, achieving an efficient dynamic intra-scale feature interaction module. The module structure is as follows: Figure 3 As shown: The first step is the self-attention interaction within the scale. This stage is consistent with the original AIFI, processing only the P5 features. The P5 features are flattened to compress the spatial dimension to 1×N, where N=40×40, resulting in sequence features. Then Simultaneously serving as Q, K, and V, global feature associations are established through multi-head self-attention, with the similarity function of self-attention employing a scaled dot product:

[0031] in To determine the dimension of the attention head, the Softmax function ensures that the attention weights are normalized. The dimension is restored from the attention output to N×C using a 1×1 step, and then restored to spatial features using a Reshape operation.

[0032] This process achieves global semantic association modeling of P5 features, capturing the spatial distribution information of defects. EDFFN is the core improvement of the module, performing frequency domain enhancement. First, a patch unrolling operation is performed. Divide the area into non-overlapping 4×4 patches, each containing 16 pixels, resulting in a patch sequence:

[0033] Where M is the total number of patches, which is 100. Then, a frequency domain transformation is performed, and a Fast Fourier Transform is applied to each patch at the pixel level to obtain the frequency domain features:

[0034] The transformation process converts spatial domain noise interference into high-frequency components, facilitating subsequent screening. Dynamic frequency domain screening is then performed, introducing a learnable quantization matrix.

[0035] Through W Weighted filtering is performed to identify low-frequency components related to defects, followed by suppression of high-frequency noise.

[0036] Where σ is the Sigmoid activation function, after processing, the filtered... Performing an inverse fast Fourier transform back to the spatial domain, we obtain:

[0037] Spatial features are then restored through patch folding:

[0038] Frequency domain enhancement is completed. Finally, the output of EDFFN is... With self-attention output The final output of the ED-AIFI module is obtained by summing the residual connections:

[0039] The output features of the encoder will provide the Decoder with a high-quality initial query.

[0040] The second submodule of the hybrid encoder is a cross-scale feature fusion module, which receives denoised high-level features and mid-to-low-level features from the feature extraction network as input. For example... Figure 4 As shown, the core of this module lies in its fusion strategy based on gated convolutional units. The original feature fusion module enhances feature representation during the training phase through reparameterization, improving efficiency during the inference phase. However, this has limitations in the photovoltaic cell defect detection scenario. Simple convolution stacking and reparameterization can easily lead to feature confusion of overlapping defects. Therefore, we replace the static convolutional block RepBlock with a gated convolutional unit ConvGLU, proposing a cross-scale feature fusion module G-CCFF. The module structure is as follows: Figure 4 As shown.

[0041] The cross-scale feature fusion module G-CCFF optimizes the fusion unit Fusion, achieving dynamic screening and accurate fusion of multiple defect features through the ConvGLU gating mechanism. The specific working principle is as follows: the input to G-CCFF consists of adjacent scale features from different stages of the backbone and the F5 values ​​processed by ED-AIFI. The high-level input features are denoted as... Low-level features are denoted as First, feature alignment is performed to ensure spatial dimension matching in the subsequent fusion process. Then, the core step of feature fusion is entered: channel adjustment is performed first. Two independent 1×1 convolutional layers are used to expand the channel dimension of the aligned features F1' and F2' respectively, increasing the number of channels from C to 2C, resulting in... and Next, ConvGLU gated fusion is performed, splitting F1'' and F2'' into value branches along the channel dimension. With gated branches The mean branch dimension is H×W×C, and the gated branch dimension is H×W×C. A 3×3 depthwise convolution with s=1 is performed on the gated branch to capture local context information, and then the gated signal is generated by the GELU activation function. The gating signal suppresses invalid features by dynamically adjusting channel weights. The value branch directly retains the initial feature information, and the final output of ConvGLU is:

[0042] Where ⊙ represents element-wise multiplication. Then, element-wise addition is performed to merge the results, resulting in the output of ConvGLU. , By adding elements together, we obtain preliminary fusion characteristics:

[0043] Finally, feature integration is performed to obtain the output. The final output is obtained by adjusting the channel dimensions to the target dimension using a 1×1 convolutional layer. The output features will serve as the input for the next level of cross-scale fusion.

[0044] Subsequently, the system automatically selects a predetermined number of feature locations from the aforementioned fused feature sequence through a query initialization module. Content queries are initialized with the feature vectors of these locations, and location queries are initialized with their spatial coordinates, collectively forming a set of high-quality initial target queries. Finally, a defect decoding module consisting of multiple Transformer decoders interacts with the initial target queries and the fused feature sequence. The decoder uses a cross-attention mechanism to repeatedly align and optimize the queries with global image features, and ultimately, the prediction head directly maps each query to a defect prediction result, including a category label and bounding box coordinates. The entire process requires no manually designed post-processing steps, achieving a completely end-to-end detection workflow.

[0045] During the training phase, a large dataset of labeled photovoltaic cell EL images is used for supervised training of all system parameters. The loss function consists of classification loss and bounding box regression loss, which is optimized using the backpropagation algorithm. Experimental results show that, while maintaining inference speed similar to mainstream real-time detection models, this system has significantly improved the detection rate of weak defects, stability in simulated noise environments, and accuracy in distinguishing densely superimposed defects on multiple public and internal EL defect test sets, effectively addressing practical pain points in the industry.

[0046] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative and not exhaustive. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A photovoltaic surface defect detection system based on image recognition processing, characterized in that, include: The feature extraction network includes multiple sequentially connected feature enhancement modules for extracting features from the input electroluminescent image and outputting a multi-scale feature map. The feature enhancement modules adopt a hybrid structure that integrates convolutional operations and additive self-attention mechanisms. A hybrid encoder, connected to the feature extraction network, is used to receive and process the multi-scale feature map, comprising: The intra-scale feature interaction module is used to perform global context modeling and noise suppression on the received high-level features; A cross-scale feature fusion module, connected to the intra-scale feature interaction module, is used to dynamically filter and fuse features from different scales through a gating mechanism and output a fused feature sequence. A decoder, connected to the hybrid encoder, is used to generate defect detection results based on the fused feature sequence.

2. The system according to claim 1, characterized in that, The feature enhancement module includes: an integration subnetwork, composed of deep convolutional layers, used for local context extraction and preliminary enhancement of input features; A convolutional additive self-attention module is used to perform attention interaction based on additive similarity calculation on the output features of the integrated subnetwork, wherein the additive similarity calculation simultaneously fuses attention information from spatial and channel dimensions; The feedforward subnetwork is used to perform a nonlinear transformation on the output of the convolutional additive self-attention module and to perform a residual connection with the output of the integration subnetwork.

3. The system according to claim 1, characterized in that, The intra-scale feature interaction module includes: Self-attention units are used to establish global spatial dependencies on the high-level feature sequences of the input; The frequency domain discriminative feedforward unit is used to perform frequency domain transformation, learnable dynamic frequency domain filtering, and inverse transformation on the output features of the self-attention unit to achieve noise suppression.

4. The system according to claim 1, characterized in that, The cross-scale feature fusion module adopts a feature fusion strategy based on gated convolutional units. It generates an independent gate signal for each input feature to perform element-level modulation on the corresponding feature values ​​and then completes the fusion.

5. A photovoltaic surface defect detection method based on image recognition processing, characterized in that, The method, applied to the system as described in any one of claims 1 to 4, comprises: The feature extraction network extracts multi-scale feature maps of the input electroluminescent image; The intra-scale feature interaction module in the hybrid encoder performs global context modeling and noise suppression on high-level features in the multi-scale feature map. The cross-scale feature fusion module in the hybrid encoder dynamically fuses noise-suppressed high-level features with features of other scales based on a gating mechanism to obtain a fused feature sequence. The decoder outputs the category and location information of the defects based on the fused feature sequence.

6. The method according to claim 5, characterized in that, In the step of extracting multi-scale feature maps through a feature extraction network, the processing procedure of the feature enhancement module includes: Perform depthwise convolution on the input features to integrate local contextual information; For the integrated features, an attention mechanism based on the additive similarity function is used to perform global interactive computation in both spatial and channel dimensions. The features that have undergone attention interaction are subjected to a nonlinear feedforward transformation and then residually connected with the input features before the depthwise convolution operation to output the enhanced features.

7. The method according to claim 5, characterized in that, The steps for global context modeling and noise suppression specifically include: Multi-head self-attention computation is performed on the high-level features to establish global feature associations; Transform the features output by the self-attention calculation to the frequency domain; Dynamically weighted filtering of frequency domain features is performed using a learnable parameter matrix to enhance the target frequency band and suppress noise frequency bands; The filtered frequency domain features are inversely converted back to the spatial domain.

8. The method according to claim 5, characterized in that, The steps for performing dynamic fusion based on the gating mechanism specifically include: For each scale of input features to be fused, generate the corresponding feature value vector and gate signal vector respectively; The eigenvalue vector is modulated element by element using the gated signal vector; The features at all scales after modulation are aggregated to obtain the fused features.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 5 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 5 to 8.

Images