Sports equipment management system based on deep learning

Abstract

The invention relates to the field of article intelligent management, in particular to a sports equipment management system based on deep learning, which comprises a real-time monitoring module, a classification identification module, a damage identification module and a user interface module, according to the invention, through carrying out cross-dimension feature fusion on the equipment room, comprehensive feature representation is realized, and accurate classification and segmentation of the equipment are realized; according to the method, a multi-head double-cross attention mechanism is adopted, multiple modal features of an equipment image are extracted, feature representations with characterization capability are provided, multiple feature representations are uniformly fused, richer and diversified feature representations are generated, and the accuracy of equipment damage detection is improved; according to the method, the most relevant features in the whole feature space are recognized and highlighted by combining enhanced self-correlation attention optimization and multi-branch combined attention mechanism processing, so that the model can consider local details of the image and also preserves the overall three-dimensional features, the feature recognition capability is improved, and the accuracy of damage classification detection is improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent management of items, specifically a sports equipment management system based on deep learning. Background Technology

[0002] With the increasing variety and quantity of equipment in stadiums, gyms, school sports facilities, and other venues, traditional manual management methods can no longer meet the needs of management efficiency and quality. Sports equipment management systems can improve management efficiency, reduce labor costs and management errors. However, traditional sports equipment management systems usually rely on manual recording of equipment borrowing and inventory information. Information updates are not timely, making it difficult to achieve real-time monitoring and management. Equipment wear and tear also depends on repeated manual inspections. Summary of the Invention

[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a deep learning-based sports equipment management system. Addressing the problem that existing sports equipment management systems typically rely on manual recording of equipment borrowing and inventory information, resulting in untimely information updates and difficulties in real-time monitoring and management, this invention employs computer vision methods to achieve rapid identification and management of equipment in the equipment room. This invention performs cross-dimensional feature fusion based on two-dimensional to three-dimensional dimensions from the equipment room's point cloud, RGB images, and depth images. It integrates depth information into the RGB images, highlighting the color and three-dimensional shape of the equipment while enabling the fusion and interaction of color and shape information. Utilizing the unique information of each modality, it achieves a more comprehensive feature representation, enabling accurate classification and segmentation of the equipment. This invention employs a multi-head dual-cross attention mechanism to extract multiple modal features from the equipment images, through pairwise... Cross-fusion enables a deeper integration of information from different modalities, providing more representative feature representations. Furthermore, multi-head cross-attention processing unifies and merges various feature representations, resulting in richer and more diverse feature representations, improving the model's expressive and generalization abilities, and increasing the accuracy of equipment damage detection. Building upon this, the present invention combines enhanced autocorrelation attention optimization with a multi-branch joint attention mechanism to extract regional features from basic image features, highlighting the hierarchical and structural relationships of features. It also captures the local structural similarity between support samples and query samples, identifying and highlighting the most relevant features in the entire feature space. This allows the model to consider both local image details and overall characteristics during damage detection, forming a comprehensive feature similarity assessment, increasing feature recognition capabilities, and improving the accuracy of damage classification and detection.

[0004] The present invention provides a sports equipment management system based on deep learning, including a real-time monitoring module, a classification and identification module, a damage identification module, and a user interface module;

[0005] The real-time monitoring module uses a point cloud scanner to scan and construct the point cloud of the sports equipment room in real time, uses a camera to collect RGB and grayscale images of the sports equipment room in real time, and uses a depth camera to collect depth images of the sports equipment room in real time.

[0006] The classification and recognition module employs a multimodal feature fusion method to extract features from the point cloud, RGB image, grayscale image, and depth image of the sports equipment room. After registering the RGB image, grayscale image, and depth image, it fuses them with the features of the point cloud to obtain three-dimensional point cloud features. Then, a 3D UNet network is used to segment the three-dimensional point cloud features to obtain a segmentation map of the equipment area. Finally, features based on a convolutional neural network are extracted from the segmentation map of the equipment area to obtain the equipment classification and recognition results.

[0007] The damage identification module employs a multi-head dual-cross attention mechanism to extract multimodal graph features from the equipment partitioning map. It then combines enhanced autocorrelation attention optimization with a multi-branch joint attention mechanism to perform damage classification and detection based on a convolutional neural network on the equipment partitioning map, thereby obtaining damage detection results for various types of equipment.

[0008] The user interface module directly controls each module and allows users to view the equipment classification and identification results and damage detection results.

[0009] Furthermore, in the damage identification module, the calculation process for the damage detection result is as follows:

[0010] Step S1: Image set partitioning, dividing all equipment partition images into support set and query set;

[0011] Step S2: Graph feature extraction. A multi-head double cross attention mechanism is used to extract graph features from the equipment partitioning map, including support graph features and query graph features.

[0012] Furthermore, step S2 specifically includes the following steps:

[0013] Step S21: Convolutional feature extraction, extracting convolutional features from the equipment partitioning map, including RGB features, depth features, grayscale features and edge features;

[0014] Step S22: Cross-attention processing. The RGB, depth, and grayscale features in the convolutional features are processed using a cross-attention function to obtain the cross-attention features: ; ; ; ; ; ;

[0015] In the formula, These represent RGB features, grayscale features, and depth features, respectively. , , , , and To preset weights, and Represents query characteristics. and Representative key features, and Representative value characteristics, Represents the activation function. and For intermediate parameters, Representative feature splicing operation, and Represents cross-features, Represents fully connected layer processing. Represents cross-attention features;

[0016] Step S23: Multi-head cross-attention processing. The cross-attention features and edge features in the convolutional features are processed using a multi-head cross-attention function to obtain the multi-head cross-attention features: ; ; ;

[0017] In the formula, Represents edge features, For intermediate parameters, , and To preset weights, Represents multi-head cross-attention characteristics;

[0018] Step S24: Multimodal cross-attention concatenation, concatenating the cross-attention features and multi-head cross-attention features and processing them with a fully connected layer to obtain graph features;

[0019] Step S3: Enhanced autocorrelation attention optimization. Perform enhanced autocorrelation attention optimization on the graph features to obtain optimized graph features: ; ;

[0020] In the formula, Representative graph features, and These represent the unfold and unsqueeze operations, respectively. Represents the Hadamard product operation. As an intermediate feature, Represents a convolutional block. These represent optimized graph features, including optimized support graph features and optimized query graph features;

[0021] Step S4: Process using a multi-branch joint attention mechanism to obtain the equipment damage classification and detection results;

[0022] Furthermore, step S4 specifically includes the following steps:

[0023] Step S41: Calculate the cross-correlation attention coefficient:

[0024] Step S411: Feature simplification. Perform 1×1 convolution on the optimized support graph features and optimized query graph features to obtain simplified support graph features and simplified query graph features.

[0025] Step S412: Feature similarity evaluation, using the cosine similarity function to calculate the cosine similarity between the simplified support graph features and the simplified query graph features: ;

[0026] In the formula, Represents simplified support graph features and simplified query graph features. Represents the cosine similarity function. Represents cosine similarity;

[0027] Step S413: Similarity convolution, performing channel expansion and channel aggregation convolution operations on cosine similarity, and using batch normalization and ReLU function for activation to obtain the convolution matching score;

[0028] Step S414: Cross-attention calculation. Construct cross-attention between support graph features and query graph features based on the convolution matching score, obtaining query cross-attention and support cross-attention: ; ;

[0029] In the formula, This represents the height and width of the supporting graph features and the query graph features. Representative at Simplified Support Graph Features of Location Features of Simplified Query Graph Match score, Represents the preset coefficient. The height-to-width ratio of the support graph features and the simplified query graph features is represented. and These represent query cross-attention and support cross-attention, respectively.

[0030] Step S415: Construct the cross-attention graph. Merge the query cross-attention and support cross-attention with the optimized support graph features and optimized query graph features respectively to obtain the query cross-attention graph and the support cross-attention graph.

[0031] Step S416: Calculate the cosine similarity between the query cross-attention map and the support cross-attention map to obtain the cross-related attention coefficient;

[0032] Step S42: Calculation of cross-soft maximum attention coefficients:

[0033] Step S421: Linear processing and weight matrix transformation. Linear layer processing is performed on the simplified support graph features and simplified query graph features. Then, the weight matrix is ​​multiplied with the simplified support graph features and simplified query graph features to obtain a new feature representation. , , and : ; ;

[0034] In the formula, , and The preset weight matrix, , , and This represents a new feature representation;

[0035] Step S422: Calculate dot product similarity and feature representation. and The dot product similarity between the two is used, and the Softmax function is applied for activation to obtain the attention distribution matrix;

[0036] Step S423: Calculate the weighted distribution features and represent the features using the attention distribution matrix. The weighted distribution characteristics are obtained by performing weighted processing.

[0037] Step S424: Calculate the weighted distribution characteristics and feature representation The cosine similarity between them is used to obtain the cross-soft-maximum attention coefficient;

[0038] Step S43: Hyperparameter classification. Multiply the cross-correlation attention coefficient and the cross-soft maximum attention coefficient by the hyperparameters α and β respectively and sum them to obtain the hyperparameter classification coefficient. Different classification coefficients correspond to different types of damage detection results.

[0039] The beneficial results achieved by the present invention using the above solution are as follows:

[0040] (1) In view of the problem that existing sports equipment management systems usually rely on manual recording of equipment borrowing and inventory information, information updates are not timely and it is difficult to achieve real-time monitoring and management, this invention adopts computer vision method to achieve rapid identification and management of equipment in the equipment room. This invention performs cross-dimensional feature fusion based on two-dimensional to three-dimensional for equipment room point cloud, RGB image and depth image, integrates depth information into RGB image, and on the basis of highlighting the color and three-dimensional shape of equipment, integrates color and shape information to achieve interaction, and uses the unique information of each modality to achieve more comprehensive feature representation, so as to achieve accurate classification and segmentation of equipment;

[0041] (2) The present invention adopts a multi-head dual cross attention mechanism to extract multiple modal features of equipment images. Through pairwise cross fusion, the information between different modalities is more deeply integrated, providing more representative feature representations. On this basis, through multi-head cross attention processing, multiple feature representations are uniformly fused to generate richer and more diverse feature representations, improve the expressive power and generalization ability of the model, and improve the accuracy of equipment damage detection.

[0042] (3) The present invention combines enhanced autocorrelation attention optimization and multi-branch joint attention mechanism to extract regional features from basic image features, highlight the hierarchical and structural relationships of features, and capture the local structural similarity between support samples and query samples on this basis. It identifies and highlights the most relevant features in the entire feature space, so that the model can consider both local image details and overall characteristics during the damage detection process, forming a comprehensive feature similarity assessment, increasing feature recognition ability, and improving the accuracy of damage classification detection. Attached Figure Description

[0043] Figure 1 A block diagram of a deep learning-based sports equipment management system provided by the present invention;

[0044] Figure 2 This is a schematic diagram of a data processing flow for a multi-headed dual-cross attention mechanism.

[0045] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0046] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0047] Example 1, see Figure 1 A deep learning-based sports equipment management system includes a real-time monitoring module, a classification and recognition module, a damage recognition module, and a user interface module.

[0048] The real-time monitoring module uses a point cloud scanner to scan and construct the point cloud of the sports equipment room in real time, uses a camera to collect RGB and grayscale images of the sports equipment room in real time, and uses a depth camera to collect depth images of the sports equipment room in real time.

[0049] The classification and recognition module employs a multimodal feature fusion method to extract features from the point cloud, RGB image, grayscale image, and depth image of the sports equipment room. After registering the RGB image, grayscale image, and depth image, it fuses them with the features of the point cloud to obtain three-dimensional point cloud features. Then, a 3D UNet network is used to segment the three-dimensional point cloud features to obtain a segmentation map of the equipment area. Finally, features based on a convolutional neural network are extracted from the segmentation map of the equipment area to obtain the equipment classification and recognition results.

[0050] The damage identification module employs a multi-head dual-cross attention mechanism to extract multimodal graph features from the equipment partitioning map. It then combines enhanced autocorrelation attention optimization with a multi-branch joint attention mechanism to perform damage classification and detection based on a convolutional neural network on the equipment partitioning map, thereby obtaining damage detection results for various types of equipment.

[0051] The user interface module directly controls each module and allows users to view the equipment classification and identification results and damage detection results.

[0052] Example 2: Based on the above examples, the calculation process for the damage detection result in the damage identification module is as follows:

[0053] Step S1: Image set partitioning, dividing all equipment partition images into support set and query set;

[0054] Step S2: Graph feature extraction. A multi-head double cross attention mechanism is used to extract graph features from the equipment partitioning map, including support graph features and query graph features.

[0055] Step S3: Enhanced autocorrelation attention optimization. Perform enhanced autocorrelation attention optimization on the graph features to obtain optimized graph features: ; ;

[0056] In the formula, Representative graph features, and These represent the unfold and unsqueeze operations, respectively. Represents the Hadamard product operation. As an intermediate feature, Represents a convolutional block. These represent optimized graph features, including optimized support graph features and optimized query graph features;

[0057] Step S4: Multi-branch joint attention mechanism processing to obtain equipment damage classification detection results.

[0058] Example 3, see Figure 2 This embodiment is based on the above embodiment. Step S2 specifically includes the following steps:

[0059] Step S21: Convolutional feature extraction, extracting convolutional features from the equipment partitioning map, including RGB features, depth features, grayscale features and edge features;

[0060] Step S22: Cross-attention processing. The RGB, depth, and grayscale features in the convolutional features are processed using a cross-attention function to obtain the cross-attention features: ; ; ; ; ; ;

[0061] In the formula, These represent RGB features, grayscale features, and depth features, respectively. , , , , and To preset weights, and Represents query characteristics. and Representative key features, and Representative value characteristics, Represents the activation function. and For intermediate parameters, Representative feature splicing operation, and Represents cross-features, Represents fully connected layer processing. Represents cross-attention features;

[0062] Step S23: Multi-head cross-attention processing. The cross-attention features and edge features in the convolutional features are processed using a multi-head cross-attention function to obtain the multi-head cross-attention features: ; ; ;

[0063] In the formula, Represents edge features, For intermediate parameters, , and To preset weights, Represents multi-head cross-attention characteristics;

[0064] Step S24: Multimodal cross-attention splicing. The cross-attention features and multi-head cross-attention features are spliced ​​and processed by a fully connected layer to obtain graph features.

[0065] Example 4: This example is based on the above examples. Step S4 specifically includes the following steps:

[0066] Step S41: Calculate the cross-correlation attention coefficient:

[0067] Step S411: Feature simplification. Perform 1×1 convolution on the optimized support graph features and optimized query graph features to obtain simplified support graph features and simplified query graph features.

[0068] Step S412: Feature similarity evaluation, using the cosine similarity function to calculate the cosine similarity between the simplified support graph features and the simplified query graph features: ;

[0069] In the formula, Represents simplified support graph features and simplified query graph features. Represents the cosine similarity function. Represents cosine similarity;

[0070] Step S413: Similarity convolution, performing channel expansion and channel aggregation convolution operations on cosine similarity, and using batch normalization and ReLU function for activation to obtain the convolution matching score;

[0071] Step S414: Cross-attention calculation. Construct cross-attention between support graph features and query graph features based on the convolution matching score, obtaining query cross-attention and support cross-attention: ; ;

[0072] In the formula, This represents the height and width of the supporting graph features and the query graph features. Representative at Simplified Support Graph Features of Location Features of Simplified Query Graph Match score, Represents the preset coefficient. The height-to-width ratio of the support graph features and the simplified query graph features is represented. and These represent query cross-attention and support cross-attention, respectively.

[0073] Step S415: Construct the cross-attention graph. Merge the query cross-attention and support cross-attention with the optimized support graph features and optimized query graph features respectively to obtain the query cross-attention graph and the support cross-attention graph.

[0074] Step S416: Calculate the cosine similarity between the query cross-attention map and the support cross-attention map to obtain the cross-related attention coefficient;

[0075] Step S42: Calculation of cross-soft maximum attention coefficients:

[0076] Step S421: Linear processing and weight matrix transformation. Linear layer processing is performed on the simplified support graph features and simplified query graph features. Then, the weight matrix is ​​multiplied with the simplified support graph features and simplified query graph features to obtain a new feature representation. , , and : ; ;

[0077] In the formula, , and The preset weight matrix, , , and This represents a new feature representation;

[0078] Step S422: Calculate dot product similarity and feature representation. and The dot product similarity between the two is used, and the Softmax function is applied for activation to obtain the attention distribution matrix;

[0079] Step S423: Calculate the weighted distribution features and represent the features using the attention distribution matrix. The weighted distribution characteristics are obtained by performing weighted processing.

[0080] Step S424: Calculate the weighted distribution characteristics and feature representation The cosine similarity between them is used to obtain the cross-soft-maximum attention coefficient;

[0081] Step S43: Hyperparameter classification. Multiply the cross-correlation attention coefficient and the cross-soft maximum attention coefficient by the hyperparameters α and β respectively and sum them to obtain the hyperparameter classification coefficient. Different classification coefficients correspond to different types of damage detection results.

[0082] Example 5: Applying this solution to the equipment management of a university sports equipment room:

[0083] The university gymnasium has a variety of sports equipment, including basketballs, volleyballs, badminton rackets, and table tennis rackets. A large number of students borrow these facilities daily. Traditional manual registration methods are inefficient and cannot monitor equipment wear and tear in a timely manner. Therefore, this system is being used for management.

[0084] System Applications:

[0085] (1) Real-time monitoring: Multiple depth cameras and point cloud scanners are installed in the equipment room to collect RGB images, depth information and three-dimensional point cloud data of the equipment in real time;

[0086] (2) Classification and recognition:

[0087] The system automatically identifies basketballs on the shelf (98.7% accuracy).

[0088] Distinguish between different models of badminton rackets (accuracy rate 97.3%);

[0089] Accurately count the inventory quantity of various equipment

[0090] (3) Damage detection:

[0091] Three basketballs were found to have severe surface wear (detection accuracy 95.2%).

[0092] Two badminton rackets were found to have broken net strings (detection accuracy rate 96.8%).

[0093] One ping-pong paddle with detached rubber was detected (detection accuracy rate 94.5%);

[0094] (4) User interface:

[0095] Administrators can view real-time inventory and damage reports via computer;

[0096] System optimization: Automatically generates repair suggestions and purchase lists;

[0097] System Implementation:

[0098] (1) Data collection:

[0099] A depth camera is used to acquire RGB and depth images, and a point cloud scanner generates a 3D point cloud of the equipment room.

[0100] (2) Classification and recognition:

[0101] By fusing multimodal features, RGB images are registered with depth images to extract 3D point cloud features;

[0102] The equipment area is segmented using a 3D UNet network, and the classification results are output (e.g., basketball area, badminton area).

[0103] (3) Damage detection:

[0104] A multi-head double cross-attention mechanism is applied to the segmented image to detect wear or air leakage on the basketball surface.

[0105] Key steps in system implementation:

[0106] Multimodal feature extraction:

[0107] RGB images are used to extract the color and texture of the equipment, depth images are used to capture its three-dimensional shape, and point cloud data is used to determine the placement position.

[0108] Applications of attention mechanisms:

[0109] Cross-attention fusion of RGB and depth features enhances the recognition of equipment labels (color coding);

[0110] Multi-head cross-focusing detection of wear and tear or air leaks in basketballs;

[0111] User interface display:

[0112] It displays a real-time equipment distribution map, marks damaged equipment, and pushes alarm information.

[0113] Example 6: Implementation details of this solution in practical application:

[0114] Hardware deployment: A combination of high-resolution depth cameras (such as Intel RealSense) and laser point cloud scanners;

[0115] Model training: Expanding damage sample data for specific equipment (such as broken badminton racket strings);

[0116] Edge computing: Lightweight models are deployed to local servers to reduce network transmission latency.

[0117] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0119] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A sports equipment management system based on deep learning, characterized in that: It includes a real-time monitoring module, a classification and identification module, a damage identification module, and a user interface module; The real-time monitoring module uses a point cloud scanner to scan and construct the point cloud of the sports equipment room in real time, uses a camera to collect RGB and grayscale images of the sports equipment room in real time, and uses a depth camera to collect depth images of the sports equipment room in real time. The classification and recognition module employs a multimodal feature fusion method to extract features from the point cloud, RGB image, grayscale image, and depth image of the sports equipment room. After registering the RGB image, grayscale image, and depth image, it fuses them with the features of the point cloud to obtain three-dimensional point cloud features. A 3D UNet network is then used to segment the three-dimensional point cloud features to obtain a segmentation map of the equipment area. Feature extraction based on a convolutional neural network is then performed on the segmentation map of the equipment area to obtain the equipment classification and recognition results. The damage identification module employs a multi-head dual-cross attention mechanism to extract multimodal graph features from the equipment partitioning map. It then combines enhanced autocorrelation attention optimization with a multi-branch joint attention mechanism to perform damage classification and detection based on a convolutional neural network on the equipment partitioning map, thereby obtaining damage detection results for various types of equipment. The user interface module directly controls each module and allows users to view the equipment classification and identification results and damage detection results.

2. The sports equipment management system based on deep learning according to claim 1, characterized in that: In the damage identification module, the calculation process for the damage detection result is as follows: Step S1: Divide all equipment partitioning maps into support sets and query sets; Step S2: Use a multi-head dual-cross attention mechanism to extract graph features of the equipment partitioning graph, including support graph features and query graph features; Step S3: Perform enhanced autocorrelation attention optimization on the graph features to obtain optimized graph features; Step S4: Multi-branch joint attention mechanism processing to obtain equipment damage classification detection results.

3. The sports equipment management system based on deep learning according to claim 2, characterized in that: Step S2 specifically includes the following steps: Step S21: Extract the convolutional features of the equipment partitioning image, including RGB features, depth features, grayscale features and edge features; Step S22: Apply a cross-attention function to the RGB features, depth features, and grayscale features in the convolutional features to obtain cross-attention features; Step S23: Process the cross-attention features and edge features in the convolutional features using a multi-head cross-attention function to obtain multi-head cross-attention features; Step S24: Concatenate the cross-attention features and multi-head cross-attention features and process them with a fully connected layer to obtain graph features.

4. A sports equipment management system based on deep learning according to claim 3, characterized in that: Step S4 specifically includes the following steps: Step S41: Calculate the cross-correlation attention coefficient; Step S42: Calculate the cross-soft maximum attention coefficient; Step S43: Hyperparameter classification. Multiply the cross-correlation attention coefficient and the cross-soft maximum attention coefficient by the hyperparameters α and β respectively and sum them to obtain the hyperparameter classification coefficient. Different classification coefficients correspond to different types of damage detection results.

5. A sports equipment management system based on deep learning according to claim 4, characterized in that: Step S41 specifically includes the following steps: Step S411: Perform 1×1 convolution on the optimized support graph features and optimized query graph features to obtain simplified support graph features and simplified query graph features; Step S412: Calculate the cosine similarity between the simplified support graph features and the simplified query graph features using the cosine similarity function; Step S413: Perform convolution operations on the cosine similarity, including channel expansion and channel aggregation, and activate it using batch normalization and the ReLU function to obtain the convolution matching score; Step S414: Construct cross-attention between support graph features and query graph features based on convolution matching scores to obtain query cross-attention and support cross-attention; Step S415: Fuse the query cross attention and support cross attention with the optimized support graph features and optimized query graph features respectively to obtain the query cross attention graph and the support cross attention graph; Step S416: Calculate the cosine similarity between the query cross-attention map and the support cross-attention map to obtain the cross-related attention coefficient.

Images