Image recognition-based adhesive quality detection method and device
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2026-02-02
- Publication Date
- 2026-05-12
Smart Images
Figure CN121616591B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of adhesive quality testing technology, specifically relating to an adhesive quality testing method and equipment based on image recognition. Background Technology
[0002] Adhesives, as an important functional material, are widely used in many core fields such as electronics, aerospace, building materials, medical and health, and automobile manufacturing. Their quality stability directly determines the bonding reliability, structural safety, and service life of the applied products. However, in the actual production process of adhesives, various factors such as raw material quality, mixing process, temperature and pressure control, and the cleanliness of the production environment can easily lead to various quality defects, including insufficient material homogeneity, particle agglomeration, gas phase inclusions, and the embedding of foreign impurities. For a long time, adhesive quality inspection has mainly relied on the traditional model of "manual visual inspection + offline sampling," which is not only costly and inefficient but also prone to missed detections. Summary of the Invention
[0003] In view of this, the present invention provides an adhesive quality inspection method and device based on image recognition, which uses computer vision technology to identify the acquired adhesive images, thereby realizing automated real-time detection of adhesive quality, and providing a new technical path for intelligent production and inspection in the adhesive industry.
[0004] The technical solution is as follows: A method for quality inspection of adhesives based on image recognition, comprising:
[0005] A digital image of the target adhesive to be inspected is acquired, and a trained recognition network is acquired. The recognition network is configured to receive the digital image and generate category information of the digital image, thereby completing the quality inspection of the target adhesive.
[0006] The recognition network includes a feature capture module, which comprises interconnected mutation capture units, semantic capture units, and adjacency fusion units. Both the mutation capture unit and the semantic capture unit have convolutional layers. The mutation capture unit is used to capture image mutation information and generate a mutation feature map. The semantic capture unit is used to capture image semantic distribution information and generate a semantic feature map. The adjacency fusion unit is used to match and fuse the mutation feature map and the semantic feature map.
[0007] In some possible implementations, the computation process of the mutation capture unit includes:
[0008] The feature map received by the mutation capture unit is subjected to a first capture process, a second capture process, and a third capture process, respectively, to generate a first feature map, a second feature map, and a third feature map. The first capture process includes a regular convolutional layer, the second capture process includes a pointwise convolutional layer, and the third capture process is used to map the feature values in the feature map to weight values.
[0009] A first transition feature is generated based on the first feature map and the second feature map;
[0010] The first feature map is subjected to a fourth capture process to generate a fourth feature map, and the second feature map is subjected to a fifth capture process to generate a fifth feature map;
[0011] A second jump feature is generated based on the fourth and fifth feature maps;
[0012] The mutation feature map is generated based on the third feature map, the first jump feature, and the second jump feature.
[0013] In some possible implementations, the above method satisfies one or more of the following conditions:
[0014] A. The first capture process includes processing using cascaded ordinary convolutional layers and activation layers;
[0015] B. The second capture process includes processing using cascaded pointwise convolutional layers and activation layers;
[0016] C. The third capture process includes processing using cascaded ordinary convolutional layers and softmax layers;
[0017] D. The fourth capture process includes processing using cascaded ordinary convolutional layers and activation layers;
[0018] E. The fifth capture process includes processing using cascaded pointwise convolutional layers and activation layers.
[0019] F. Generating a first jump feature based on the first feature map and the second feature map includes: subtracting the corresponding elements of the first feature map and the second feature map, and then activating them with the Tanh function to generate the first jump feature;
[0020] G. Generating a second jump feature based on the fourth and fifth feature maps, including: subtracting the corresponding elements of the fourth and fifth feature maps and then activating them with the Tanh function to generate the second jump feature;
[0021] H. Generating the mutation feature map based on the third feature map, the first jump feature, and the second jump feature includes: multiplying the corresponding elements of the third feature map, the first jump feature, and the second jump feature and then activating them to generate the mutation feature map.
[0022] In some possible implementations, the computation process of the semantic capture unit includes:
[0023] The spatial distribution features of the first transition feature and the second transition feature are extracted respectively to generate the corresponding first distribution feature and second distribution feature;
[0024] The fourth feature map is multiplied with the corresponding element of the second distribution feature to generate the first structural feature; the fifth feature map is multiplied with the corresponding element of the first distribution feature to generate the second structural feature.
[0025] Semantic capture processing is performed on the first and second structural features respectively to generate corresponding third and fourth structural features;
[0026] The third structural feature is fused with the fourth structural feature to generate the semantic feature map.
[0027] In some possible implementations, the above method satisfies one or more of the following conditions:
[0028] I. Extract the spatial distribution features of the first transition feature and the second transition feature respectively, including: processing the first transition feature and the second transition feature respectively using a series of channel global pooling layers and activation layers;
[0029] J. The semantic capture process includes processing using cascaded dimensionality-reducing convolutional layers and activation layers;
[0030] K. The fusion of the third structural feature and the fourth structural feature includes: after the third structural feature and the fourth structural feature are concatenated, they are processed by a series of ordinary convolutional layers and activation layers.
[0031] In some possible implementations, the calculation process of the adjacency fusion unit includes:
[0032] A first adjacency feature is generated based on the first distribution feature and the second distribution feature;
[0033] The first adjacency feature is activated by multiplying the corresponding element of the mutation feature map to generate the second adjacency feature;
[0034] After the second adjacency feature is pooled, it is added to the corresponding element of the semantic feature map to generate a matching feature map, thus completing the matching and fusion of the mutation feature map and the semantic feature map.
[0035] In some possible implementations, the first distribution feature and the corresponding elements of the second distribution feature are added together to activate the first adjacent feature.
[0036] In some possible implementations, the recognition network further includes a classification module that generates category information of the digital image based on the fusion results of the adjacency fusion unit.
[0037] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method described above.
[0038] Compared with the prior art, the beneficial effects of the present invention are:
[0039] In actual production scenarios, adhesives may exhibit single-type quality defects or multiple defects superimposed on each other. Defects may be localized in a small area or distributed over a large area. The visual manifestations of adhesive defects are highly diverse, and existing image recognition models often struggle to simultaneously capture both local and global information. They are prone to misjudging small-scale defects or sparsely distributed defects. This invention incorporates a mutation capture unit for capturing image mutation information and a semantic capture unit for capturing image semantic distribution information within the feature capture module. Furthermore, an adjacency fusion unit matches and fuses the mutation feature map with the semantic feature map. This allows the recognition network to simultaneously capture local mutation defects (such as bubbles, agglomerations, and impurity particles, manifested as rapid changes in local pixels) and overall semantic distribution defects (such as gas phase inclusions, impurity embeddings, and color deviations, manifested as abnormal global pixel distribution patterns), significantly improving the accuracy and reliability of adhesive quality identification. Attached Figure Description
[0040] Figure 1 This is a schematic diagram showing the connection relationship between multiple feature capture modules in this application.
[0041] In the attached image:
[0042] 1-Feature capture module, 2-Mutation capture unit, 3-Semantic capture unit, 4-Adjacency fusion unit. Detailed Implementation
[0043] The embodiments of the present invention will now be described with reference to the accompanying drawings. The terminology used in the embodiments section is for illustrative purposes only and is not intended to limit the scope of the invention.
[0044] The processing order described in the following embodiments is merely an example and is not limited to the processing order set forth herein. Rather, the processing order may be changed based on the implementation methods disclosed in this application, except for processes that must occur in a specific order. All other embodiments obtained through these changes are within the scope of protection of this application.
[0045] To address the technical problems of high quality cost, low efficiency, and low reliability in current adhesive testing methods, this invention provides an image recognition-based adhesive quality testing method. This method utilizes a trained image recognition network to classify and identify acquired adhesive images, thereby achieving automated and rapid quality testing of adhesives.
[0046] Existing image recognition models struggle to simultaneously consider both local and global information during image feature extraction, resulting in a high misjudgment rate for small-scale defects and sparsely distributed overall defects in adhesives. Therefore, in some embodiments, the image recognition-based adhesive quality inspection method provided by this invention includes the following steps: acquiring a digital image of the target adhesive to be inspected; acquiring a trained recognition network; the recognition network is configured to receive the digital image and generate category information for the digital image, thereby completing the quality inspection of the target adhesive.
[0047] The recognition network includes a feature capture module 1, which comprises interconnected mutation capture unit 2, semantic capture unit 3, and adjacent fusion unit 4. Both mutation capture unit 2 and semantic capture unit 3 have convolutional layers. Mutation capture unit 2 is used to capture image mutation information and generate mutation feature maps. Semantic capture unit 3 is used to capture image semantic distribution information and generate semantic feature maps. Adjacency fusion unit 4 is used to match and fuse the mutation feature maps and semantic feature maps to generate matching feature maps.
[0048] Image mutation information refers to the feature information carried by regions in a digital image where pixel value gradients (rate of change of brightness / color / grayscale) rise or fall sharply, reflecting significant differences in pixel features (grayscale, color, texture, etc.) between this region and adjacent regions. Image semantic distribution information refers to the distribution pattern, statistical characteristics, and spatial correlation of pixel features (grayscale, color, texture, and brightness, etc.) in a digital image on a global scale, reflecting the structural pattern, uniformity, and consistency of the entire image. By matching and fusing image mutation information with image semantic distribution information, both details and the overall picture are considered, greatly improving the accuracy and reliability of the recognition network in identifying adhesive quality.
[0049] In some implementations, for adhesives with high transparency, digital images can be obtained by directly photographing the adhesive in the exposure environment. For adhesives with low transparency, since internal defects are obscured and difficult to observe, digital images of the adhesive can be obtained by sampling and flattening it at different locations on the product before photographing. The training process and method of the above-described recognition network are consistent with the training process and method of existing recognition models. During training, cross-entropy can be used as the loss function, and the recognition network training is considered complete when the loss function converges.
[0050] In some implementations, the identification network includes multiple sequentially connected feature capture modules. For example, there may be 2, 3, 4, 5, 6, or 7 feature capture modules. In this case, the first feature capture module is configured to receive a digital image of the adhesive. The matching feature map generated by the adjacency fusion unit inside the first feature capture module is used as the information output by the feature capture module to the second feature capture module, and so on. The matching feature map generated by the adjacency fusion unit inside the Nth feature capture module is used as the information output by the feature capture module to the N+1th feature capture module.
[0051] The category information of network production can be represented by different numbers. For example, the numbers 1-40 can represent different categories. If the identified category information is 1-10, it means that most of the defects in the adhesive are due to insufficient material homogeneity. The larger the number, the more defects the adhesive has and the worse its quality. Similarly, if the identified category information is 11-20, it means that most of the defects in the adhesive are due to particle agglomeration. The larger the number, the more defects the adhesive has and the worse its quality. If the identified category information is 21-30, it means that most of the defects in the adhesive are due to gas phase inclusions. The larger the number, the more defects the adhesive has and the worse its quality. If the identified category information is 31-40, it means that most of the defects in the adhesive are due to the embedding of foreign impurities. The larger the number, the more defects the adhesive has and the worse its quality. This approach allows for the identification of major defect types and the overall quality of products under various defect conditions. For example, categories 1, 11, 21, or 31 indicate that the adhesive has few or almost no defects overall. This makes it easy for users to quickly identify the cause of the problem based on the identification results and to flexibly set the inspection pass standards according to the actual needs of the product.
[0052] The mutation capture unit 2, semantic capture unit 3, and adjacency fusion unit 4 described above can each be implemented independently using existing algorithms capable of achieving their respective functions. In some possible implementations, the calculation process of mutation capture unit 2 includes:
[0053] S101. Perform first capture processing, second capture processing, and third capture processing on the feature maps received by the mutation capture unit, respectively, to generate first feature maps, second feature maps, and third feature maps. The first capture processing includes a regular convolutional layer, the second capture processing includes a pointwise convolutional layer, and the third capture processing is used to map the feature values in the feature maps to weight values. For example, the first capture processing includes processing using a concatenated first convolutional layer and a first activation layer; the second capture processing includes processing using a concatenated first pointwise convolutional layer and a second activation layer; and the third capture processing includes processing using a concatenated second convolutional layer and a softmax layer.
[0054] S102. Generate a first transition feature based on the first feature map and the second feature map. For example, subtract the corresponding element in the second feature map from the element in the first feature map, and then activate it through the first Tanh function to generate the first transition feature.
[0055] S103. Perform a fourth capture process on the first feature map to generate a fourth feature map, and perform a fifth capture process on the second feature map to generate a fifth feature map. For example, the fourth capture process includes processing using a cascaded third convolutional layer and a third activation layer; the fifth capture process includes processing using a cascaded second pointwise convolutional layer and a fourth activation layer.
[0056] S104. Generate the second transition feature based on the fourth and fifth feature maps. For example, subtract the corresponding element in the fifth feature map from the element in the fourth feature map, and then activate it using the second Tanh function to generate the second transition feature.
[0057] S105. Generate a mutation feature map based on the third feature map, the first jump feature, and the second jump feature. For example, the corresponding elements of the third feature map, the first jump feature, and the second jump feature are multiplied together and then activated by the fifth activation layer to generate the mutation feature map.
[0058] The first and second pointwise convolutional layers described above both have a 1×1 kernel size and a stride of 1, performing convolution operations independently on each spatial location on the feature map. The first, second, and third convolutional layers are all ordinary convolutional layers, and their kernels are not particularly limited. In some preferred embodiments, the kernel size of the first, second, and third convolutional layers is 3×3, and the stride is 1. The first, second, third, fourth, and fifth activation layers described above are not particularly limited and can use existing conventional activation functions. In some preferred embodiments, the first, second, third, and fourth activation layers all use the ReLU function, and the fifth activation layer uses the sigmoid function.
[0059] The mutation capture unit 2 first extracts the abrupt change features from the image, and then captures the mutation features from the image through the abrupt change features. This two-step progressive change feature capture method has a more comprehensive feature coverage capability, and can better cope with the complex and diverse defect morphologies presented by adhesives. Furthermore, in the process of generating mutation features through abrupt change features, this application creatively introduces a third feature map. The weight values in the third feature map can selectively purify and regulate the abrupt change features, which can effectively avoid the recognition network being interfered with by false defect signals, and prevent the network from being overly influenced by local high-frequency information and neglecting the attention to global features, thus improving the robustness and accuracy of the model.
[0060] In some possible implementations, based on the mutation capture unit 2 described above, the calculation process of the semantic capture unit 3 in the same feature capture module 1 includes:
[0061] S201. Extract the spatial distribution features of the first and second jump features respectively, and generate the corresponding first and second distribution features. For example, process the first jump feature using a cascaded channel global pooling layer and a sixth activation layer to obtain the first distribution feature, and process the second jump feature using a cascaded channel global pooling layer and a seventh activation layer to obtain the second distribution feature. The aforementioned channel global pooling layer can be an existing channel global max pooling layer, which extracts the maximum eigenvalue at different spatial locations of the feature map to obtain a single-channel matrix. For example, for a feature map with height × width × channels of H × W × C, it contains H × W vectors of length C with different spatial locations. When the channel global max pooling layer processes the feature map, it extracts the maximum eigenvalue of each vector, and then uses the extracted eigenvalue to output a matrix of size H × W × 1.
[0062] S202. The fourth feature map is multiplied by the corresponding elements of the second distribution feature to generate the first structural feature; the fifth feature map is multiplied by the corresponding elements of the first distribution feature to generate the second structural feature. The width and height of the second distribution feature are equal to the width and height of the fourth feature map, and the width and height of the first distribution feature are equal to the width and height of the fifth feature map. Multiplying the fourth feature map by the corresponding elements of the second distribution feature means that each layer of the fourth feature map (each channel of the feature map corresponds to one layer) is multiplied by the corresponding elements of the second distribution feature. Similarly, multiplying the fifth feature map by the corresponding elements of the first distribution feature means that each layer of the fifth feature map (each channel of the feature map corresponds to one layer) is multiplied by the corresponding elements of the first distribution feature.
[0063] S203. Semantic capture processing is performed on the first structural feature and the second structural feature respectively to generate corresponding third structural feature and fourth structural feature. For example, in some embodiments, a cascaded fourth convolutional layer and an eighth activation layer can be used to perform semantic capture processing on the first structural feature, and a cascaded fifth convolutional layer and a ninth activation layer can be used to perform semantic capture processing on the second structural feature. As a further, non-limiting example, a fourth convolutional layer and a fifth convolutional layer with a kernel size of 3×3 and a stride of 2 can be used to process the first structural feature and the second structural feature respectively, thereby achieving the purpose of reducing the spatial dimension of the first structural feature and the second structural feature.
[0064] S204. The third and fourth structural features are fused to generate a semantic feature map. For example, after the third and fourth structural features are concatenated, they are processed by a cascaded fifth convolutional layer and a ninth activation layer. As a further, non-restrictive example, the kernel size of the fifth convolutional layer can be 7×7, and the stride can be 1.
[0065] The sixth, seventh, eighth, and ninth activation layers are not particularly limited and can be processed using existing conventional activation functions. In some preferred embodiments, the sixth and seventh activation layers both use the sigmoid function, and the eighth and ninth activation layers both use the ReLU function.
[0066] When the semantic capture unit 3 adopts the above technical solution, it performs correlation deduction of the semantic information and spatial semantic distribution law of the adhesive image based on the joint feature map of the jump features of multiple local defects, realizing the linkage mining of local semantic features and global distribution features, and strengthening the recognition network's ability to understand the semantics of different image features.
[0067] In some possible implementations, based on the above-described mutation capture unit 2 and semantic capture unit 3, the calculation process of the adjacency fusion unit 4 in the same feature capture module 1 includes:
[0068] S301. Generate a first adjacency feature based on the first distribution feature and the second distribution feature. For example, the first distribution feature and the second distribution feature have the same size, and the corresponding elements of the first distribution feature and the second distribution feature are added together and activated (e.g., using the ReLU function) to generate the first adjacency feature.
[0069] S302. The first adjacent feature is activated by multiplying the corresponding element of the mutation feature map (e.g., using the ReLU function) to generate the second adjacent feature. Specifically, the width and height of the first adjacent feature are equal to the width and height of the mutation feature map. The mutation feature map is multiplied by the corresponding element of the first adjacent feature, that is, each layer of the mutation feature map is multiplied by the corresponding element of the first adjacent feature.
[0070] S303. After pooling, the second adjacent feature is added to the corresponding element of the semantic feature map to generate a matching feature map. This completes the matching and fusion of the mutation feature map and the semantic feature map. The matching feature map serves as the output feature map of this feature capture module. In the next calculation, this matching feature map is used as the feature map received by the mutation capture unit in the next feature capture module. For example, the spatial dimension of the second adjacent feature can be reduced by performing existing local average pooling. The size of the pooling window and the step size of the pooling operation during local average pooling can be set according to the actual situation (for example, the pooling window size can be 2×2, and the pooling operation step size can be 2), as long as the feature obtained after pooling is equal in size to the semantic feature map in this feature capture module.
[0071] When the adjacency fusion unit adopts the above scheme, it uses the first adjacency feature as a bridge for the interaction between adjacency semantic distribution information and image mutation information, so as to realize the contextual transitional fusion of semantic distribution information and image mutation information, improve the feature fusion effect, and thus improve the defect detection rate and accuracy.
[0072] Based on the above configuration, such as Figure 1 As shown, the feature map received by the mutation capture unit 2 inside the first feature capture module 1 is the digital image of the adhesive. The feature map received by the mutation capture unit 2 inside the second feature capture module 1 is the matching feature map generated by the adjacent fusion unit 4 inside the first feature capture module 1. And so on, the feature map received by the mutation capture unit 2 inside the N+1th feature capture module 1 is the matching feature map generated by the adjacent fusion unit 4 inside the Nth feature capture module 1.
[0073] In some possible implementations, the recognition network further includes a classification module configured to receive the matching feature map output by the last feature capture module and generate category information for the digital image. This application does not impose any particular limitation on the classification module; it can be implemented using numerous existing technologies. For example, as one exemplary implementation, the classification module includes a spatial global pooling layer, a fully connected layer, and a classifier (e.g., a softmax classifier) arranged in series. The aforementioned spatial global pooling layer can be a spatial global average pooling layer, which calculates the average value of the feature values in each layer of the feature map (each layer corresponds to one channel) to obtain a one-dimensional feature vector. The classifier calculates and outputs category information.
[0074] Another aspect of this application provides an electronic device, including: a memory for storing one or more programs; and a controller; when the one or more programs are executed by the controller, the image recognition method described above is performed. As an exemplary electronic device, it may include a central processing unit (CPU), which can perform calculations based on a program stored in read-only memory (ROM) to complete the method in the above embodiments, or perform calculations based on a program stored in random access memory (RAM) to complete the method in the above embodiments.
[0075] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention. Those skilled in the art, under the guidance of the present invention, can make various similar representations without departing from the spirit and claims of the present invention, and such modifications all fall within the protection scope of the present invention.
Claims
1. A method for quality inspection of adhesives based on image recognition, characterized in that, Includes the following steps: A digital image of the target adhesive to be inspected is acquired, and a trained recognition network is acquired. The recognition network is configured to receive the digital image and generate category information of the digital image, thereby completing the quality inspection of the target adhesive. The recognition network includes a feature capture module, which comprises interconnected mutation capture units, semantic capture units, and adjacency fusion units. Both the mutation capture unit and the semantic capture unit have convolutional layers. The mutation capture unit is used to capture image mutation information and generate mutation feature maps, and the semantic capture unit is used to capture image semantic distribution information and generate semantic feature maps. The adjacency fusion unit is used to match and fuse the mutation feature maps and the semantic feature maps. The calculation process of the mutation capture unit includes: The feature map received by the mutation capture unit is subjected to a first capture process, a second capture process, and a third capture process, respectively, to generate a first feature map, a second feature map, and a third feature map. The first capture process includes a regular convolutional layer, the second capture process includes a pointwise convolutional layer, and the third capture process is used to map the feature values in the feature map to weight values. A first transition feature is generated based on the first feature map and the second feature map; The first feature map is subjected to a fourth capture process to generate a fourth feature map, and the second feature map is subjected to a fifth capture process to generate a fifth feature map; A second jump feature is generated based on the fourth and fifth feature maps; The mutation feature map is generated based on the third feature map, the first jump feature, and the second jump feature; The calculation process of the semantic capture unit includes: The spatial distribution features of the first transition feature and the second transition feature are extracted respectively to generate the corresponding first distribution feature and second distribution feature; The fourth feature map is multiplied with the corresponding element of the second distribution feature to generate the first structural feature; the fifth feature map is multiplied with the corresponding element of the first distribution feature to generate the second structural feature. Semantic capture processing is performed on the first and second structural features respectively to generate corresponding third and fourth structural features; The third structural feature is fused with the fourth structural feature to generate the semantic feature map.
2. The method according to claim 1, characterized in that, One or more of the following conditions must be met: A. The first capture process includes processing using cascaded ordinary convolutional layers and activation layers; B. The second capture process includes processing using cascaded pointwise convolutional layers and activation layers; C. The third capture process includes processing using cascaded ordinary convolutional layers and softmax layers; D. The fourth capture process includes processing using cascaded ordinary convolutional layers and activation layers; E. The fifth capture process includes processing using cascaded pointwise convolutional layers and activation layers.
3. The method according to claim 1, characterized in that, One or more of the following conditions must be met: F. Generating a first jump feature based on the first feature map and the second feature map includes: subtracting the corresponding elements of the first feature map and the second feature map, and then activating them with the Tanh function to generate the first jump feature; G. Generating a second jump feature based on the fourth and fifth feature maps, including: subtracting the corresponding elements of the fourth and fifth feature maps and then activating them with the Tanh function to generate the second jump feature; H. Generating the mutation feature map based on the third feature map, the first jump feature, and the second jump feature includes: multiplying the corresponding elements of the third feature map, the first jump feature, and the second jump feature and then activating them to generate the mutation feature map.
4. The method according to claim 1, characterized in that, One or more of the following conditions must be met: I. Extract the spatial distribution features of the first transition feature and the second transition feature respectively, including: processing the first transition feature and the second transition feature respectively using a series of channel global pooling layers and activation layers; J. The semantic capture process includes processing using cascaded dimensionality-reducing convolutional layers and activation layers; K. The fusion of the third structural feature and the fourth structural feature includes: after the third structural feature and the fourth structural feature are concatenated, they are processed by a series of ordinary convolutional layers and activation layers.
5. The method according to claim 1, characterized in that, The calculation process of the adjacency fusion unit includes: A first adjacency feature is generated based on the first distribution feature and the second distribution feature; The first adjacency feature is activated by multiplying the corresponding element of the mutation feature map to generate the second adjacency feature; After the second adjacency feature is pooled, it is added to the corresponding element of the semantic feature map to generate a matching feature map, thus completing the matching and fusion of the mutation feature map and the semantic feature map.
6. The method according to claim 5, characterized in that, The first distribution feature and the second distribution feature are added together to activate the first adjacent feature.
7. The method according to claim 1, characterized in that, The recognition network also includes a classification module, which generates category information of the digital image based on the matching and fusion results of the adjacency fusion unit.
8. An electronic 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, it implements the method as described in any one of claims 1 to 7.