Profile Image Target Recognition Method Based on Cross-Attention Guided Feature Fusion

Abstract

This invention proposes a profile image target recognition method based on cross-attention guided feature fusion. The method includes: selecting 3D models from a dataset and randomly combining them into an indoor scene model using modeling software; sectioning the indoor scene model along the same axis to obtain an ASD profile image dataset; inputting the ASD profile image dataset into a dual-input backbone network for feature extraction to obtain branch-extracted feature maps; performing cross-attention guided feature fusion on the branch-extracted feature maps to obtain a fused feature map; performing multi-scale feature enhancement on the fused feature map to obtain an enhanced multi-scale feature map; and passing the enhanced multi-scale feature map through the neck and head sections to obtain the target recognition result. This invention enhances the multi-scale feature representation capability through cascaded groups and cross-global attention feature mechanisms, and significantly reduces parameter redundancy by employing an adaptive shared convolutional head, thereby improving the accuracy and efficiency of profile image recognition.

Description

Technical Field

[0001] This invention relates to the field of computer vision target recognition technology, and in particular to a profile target recognition method based on cross-attention guided feature fusion. Background Technology

[0002] Cross-sectional views are a crucial technology in civil engineering for revealing internal structures and spatial organization. However, target identification in unlabeled cross-sectional views still faces two fundamental challenges: similar objects may exhibit significant shape differences at different cutting locations, while dissimilar objects may present similar forms on different cutting planes. This characteristic significantly increases the difficulty of identification. For a set of cross-sectional slice images of the same object along the same axis, although they provide rich information, how to effectively integrate this information becomes a key issue.

[0003] Object recognition, as a core task of computer vision, plays a crucial role in numerous fields such as medical image analysis and autonomous driving perception systems. With the development of deep learning technology, multimodal detection methods have gradually become a research hotspot. By fusing data sources from different modalities, they effectively solve the problem of complex scene recognition that is difficult to handle with a single modality. Although existing multimodal object detection algorithms have made significant progress in feature fusion strategies, they usually rely on simple interaction mechanisms, making it difficult to fully explore the deep semantic relationships between profile images. Summary of the Invention

[0004] In view of the above, the main objective of this invention is to propose a profile map target recognition method based on cross-attention guided feature fusion to solve the aforementioned technical problems.

[0005] This invention proposes a profile image target recognition method based on cross-attention guided feature fusion, the method comprising the following steps:

[0006] Step 1: Select 3D models from the dataset and randomly combine them into an indoor scene model using modeling software; cut the indoor scene model along the same axis to obtain the ASD profile dataset.

[0007] Step 2: Input the ASD profile dataset into the dual-input backbone network to extract features from two consecutive profiles in parallel, resulting in a dual-branch extracted feature map.

[0008] Step 3: Perform cross-attention guided feature fusion on the feature maps extracted from the two branches to fuse the complementary information of the two profile maps and obtain the fused feature map.

[0009] Step 4: Perform multi-scale feature enhancement processing on the fused feature map to obtain the enhanced multi-scale feature map;

[0010] Step 5: After the enhanced multi-scale feature map passes through the neck and head, the target recognition result is obtained.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0012] 1. When performing target recognition tasks on building cross-section diagrams, this invention can effectively fuse complementary information from two cross-section diagrams through the Cross Attention Guided Feature Fusion (CAG-Fusion) module, effectively solving the problems of previous methods that could not fully utilize the information from multiple cross-section diagrams and lacked feature representation capabilities.

[0013] 2. This invention enhances the multi-scale feature representation capability through cascaded groups and cross-global attention feature mechanism (CGCAFM), and significantly reduces parameter redundancy by adopting adaptive shared convolutional head (ASC-Head), thereby improving the accuracy and efficiency of profile image recognition. Attached Figure Description

[0014] Figure 1 This is a flowchart of the profile target recognition method based on cross-attention guided feature fusion proposed in this invention;

[0015] Figure 2 This is a schematic diagram illustrating the construction of the ASD dataset for the profile map target recognition method based on cross-attention guided feature fusion proposed in this invention;

[0016] Figure 3 This is a schematic diagram of the dual-input backbone network architecture of the profile map target recognition method based on cross-attention guided feature fusion proposed in this invention;

[0017] Figure 4 This is a schematic diagram of the CAG-Fusion module structure of the profile map target recognition method based on cross-attention guided feature fusion proposed in this invention;

[0018] Figure 5 This is a schematic diagram of the CGCAFM module structure of the profile map target recognition method based on cross-attention guided feature fusion proposed in this invention;

[0019] Figure 6 This is a schematic diagram of the ASC-Head module structure of the profile map target recognition method based on cross-attention guided feature fusion proposed in this invention;

[0020] Figure 7 This diagram illustrates a comparison of the test results of the profile map object recognition model (SF-YOLO), the multimodal fusion object recognition model CFT, and the YOLO11 model on the ASD dataset, based on cross-attention guided feature fusion. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0023] Please see Figure 1 This invention proposes a profile target recognition method based on cross-attention guided feature fusion, which includes the following steps:

[0024] Step 1: Select 3D models from the dataset and randomly combine them into an indoor scene model using modeling software; cut the indoor scene model along the same axis to obtain the ASD profile dataset.

[0025] Furthermore, in step 1, five common furniture 3D models were selected from the ShapeNet dataset, with each type of furniture having multiple different styles. These models were then randomly combined into 15 sets of interior scene models using SketchUp 2023 to generate cross-sectional views.

[0026] To ensure spatial consistency between the two sections, a pair of registered section images is generated by cutting along the same axis, and the cutting spacing is controlled within the range of 5% to 20% of the model size to balance complementarity and correlation.

[0027] The final ASD profile dataset contains 1,200 pairs of registered profile views and corresponding annotation information, which are used to verify the effectiveness of the algorithm.

[0028] For details, please refer to Figure 2 , Figure 2 This is a diagram illustrating the process of creating the dataset. Figure 2 In this context, X, Y, and Z represent the three principal axes of the 3D model. Figure 2 The numbers 1, 2, 3, ... N marked with circles represent the indices of the cutting planes arranged sequentially along that axis. Figure 2 In this context, N represents the number of the maximum cutting plane in that direction;

[0029] The top row of the right column displays the cross-sectional images extracted from the corresponding cut sections, while the bottom row shows the corresponding cross-sectional recognition results obtained using this invention.

[0030] Step 2: Input the ASD profile dataset into the dual-input backbone network to extract features from two consecutive profiles in parallel, resulting in a dual-branch extracted feature map.

[0031] In step 2, the ASD profile dataset is input into a dual-input backbone network to extract features from two consecutive profiles in parallel, resulting in a dual-branch extracted feature map. The corresponding process has the following relationship:

[0032] ;

[0033] in, This represents the feature map extracted by the first branch. This represents the feature map extracted by the second branch. This indicates that the feature extraction process is performed through the first branch. This indicates that feature extraction is performed through the second branch. This represents the input to the first branch. This represents the input for the second branch.

[0034] For further details, please refer to Figure 3 In this step, the dual-input backbone network architecture simultaneously extracts features from two registration profiles. Each branch is built based on the YOLO11 architecture, which mainly consists of three parts: the backbone network, the neck network, and the detection head. Figure 3 The top left corner shows the input cross-sectional view. Figure 3 The bottom right corner shows the recognition results output by the network. Figure 3 The upper right corner contains the basic module units that make up the network.

[0035] Step 3: Perform cross-attention guided feature fusion on the feature maps extracted from the two branches to fuse the complementary information of the two profile maps and obtain the fused feature map.

[0036] Please see Figure 4 In step 3, cross-attention guided feature fusion is performed on the feature maps extracted from the two branches to fuse the complementary information of the two profile maps, resulting in a fused feature map. The specific steps are as follows:

[0037] An adaptive transformation function is used to align the feature maps extracted from the branches by channel dimensions, resulting in a profile feature map after adaptive feature alignment.

[0038] After the profile feature map after adaptive feature alignment and the feature map extracted by the branch are concatenated along the channel dimension, an adaptive weighting process is then applied to obtain the feature map weighted by the attention mechanism.

[0039] By using the feature map weighted by the attention mechanism, the feature map extracted by the branch, and the ASD profile map, and through the cross-branch information interaction mechanism, the cross-section view feature map after cross-branch residual enhancement is obtained.

[0040] Using the cross-branch residual enhanced section view feature map, interactive features are obtained through element-wise multiplication.

[0041] The interaction features are processed sequentially through convolutional projection, batch normalization, and sigmoid activation to obtain context-guided weights.

[0042] By using context-guided weights, the cross-sectional view feature map after cross-branch residual enhancement is stitched together along the channel dimension to obtain the fused feature map.

[0043] In the process of using an adaptive transformation function to align the channel dimensions of the feature maps extracted from the branches, resulting in the adaptively aligned profile feature map, the following relationship exists:

[0044] ;

[0045] in, This represents the profile feature map after adaptive feature alignment. This indicates that the process is handled through an adaptive transformation function;

[0046] Furthermore, the adaptive transformation function in When, channel alignment is achieved through a 1×1 convolutional layer; when When, it is an identity mapping. This represents the number of channels in the feature map extracted from the branch. This indicates the number of channels in the feature map extracted by the second branch;

[0047] After concatenating the profile feature map (adaptively aligned) and the feature map extracted from the branches along the channel dimension, and then performing adaptive weighting to obtain the attention-weighted feature map, the following relationship exists:

[0048] ;

[0049] in, This represents the feature map after being weighted by the attention mechanism. This indicates that the process has undergone adaptive weighting. This indicates that the data has undergone splicing processing along the channel dimension;

[0050] In the process of using the attention-weighted feature map, the branch-extracted feature map, and the ASD profile map, and processing them through a cross-branch information interaction mechanism to obtain the cross-section view feature map after cross-branch residual enhancement, the following relationship exists:

[0051] ;

[0052] in, This represents the first cross-sectional view feature map after cross-branch residual enhancement. This represents the first cross-sectional view feature map after cross-branch residual enhancement. This represents the feature map in the first branch after being weighted by the attention mechanism. This represents the feature map in the second branch after being weighted by the attention mechanism. This represents element-wise addition.

[0053] In the process of obtaining interactive features by using the cross-branch residual-enhanced section view feature map and performing element-wise multiplication, the following relationship exists:

[0054] ;

[0055] in, Indicates interactive features, This represents element-wise multiplication;

[0056] In the process of processing the interaction features sequentially through convolutional projection, batch normalization, and sigmoid activation to obtain the context-guided weights, the following relationship exists:

[0057] ;

[0058] in, Indicates context-guided weights, This indicates that the signal has been processed using the sigmoid activation function. This indicates that the batch normalization process has been performed. This indicates that the image has undergone 1×1 convolutional projection processing.

[0059] In the process of using context-guided weights to perform channel-dimensional stitching on the cross-branch residual enhancement feature map to obtain the fused feature map, the following relationship exists:

[0060] ;

[0061] in, This represents the fused feature map.

[0062] Furthermore, through this process, efficient fusion of cross-section semantics has been achieved.

[0063] Step 4: Perform multi-scale feature enhancement processing on the fused feature map to obtain the enhanced multi-scale feature map;

[0064] Please see Figure 5 In step 4, multi-scale feature enhancement processing is performed on the fused feature map to obtain the enhanced multi-scale feature map. The specific steps are as follows:

[0065] This step combines the cascaded attention mechanism and the convolution-attention feature fusion mechanism;

[0066] The fused feature maps are divided into groups, and 1×1 convolutions are used to jointly extract the feature maps in each group to obtain query vectors, key vectors, and value vectors.

[0067] Attention is calculated using query vectors, key vectors, and value vectors to obtain the output of the global attention branch;

[0068] The query vector, key vector, and value vector are subjected to depthwise separable convolution to obtain the enhanced query vector, enhanced key vector, and enhanced value vector, respectively.

[0069] By integrating the enhanced query vector, enhanced key vector, and enhanced value vector through a fully connected layer, the output of the local convolutional branch is obtained.

[0070] Based on the output of the global attention branch and the output of the local convolution branch, through Convolution processing yields output features fused with convolutional and attention features.

[0071] The output features of the convolution-attention feature fusion are used as group outputs. The current group output is added to the feature map in the next group element by element to obtain the output features processed by the cascaded group attention mechanism.

[0072] After concatenating all the output features processed by the cascaded attention mechanism along the channel dimension, a linear projection is performed to obtain the final output features of the cascaded attention mechanism.

[0073] Utilizing the final output features of the cascaded attention mechanism and a given feature map, through linear transformation and... After convolution processing, an enhanced multi-scale feature map is obtained;

[0074] In the process of jointly extracting the feature maps from each group using 1×1 convolution to obtain the query vector, key vector, and value vector, the following relationship exists:

[0075] ;

[0076] in, Represents the query vector. Represents the key vector. Represents a value vector. Indicates passage Convolutional joint extraction processing, Indicates the first Feature maps in each group;

[0077] In the process of calculating attention using query vectors, key vectors, and value vectors to obtain the output of the global attention branch, the following relationship exists:

[0078] ;

[0079] in, This represents the output of the global attention branch. This represents the transpose of the key vector. This represents the temperature parameter of the science department. This represents the number of channels for each self-attention head;

[0080] In the process of integrating information through a fully connected layer using the enhanced query vector, enhanced key vector, and enhanced value vector to obtain the output of the local convolutional branch, the following relationship exists:

[0081] ;

[0082] in, This represents the output of a local convolution branch. This indicates that the data has undergone a 3×3×3 depthwise convolution. This indicates that the process has been handled by a fully connected layer. This indicates a feature concatenation operation. This represents the enhanced query vector. This represents the enhanced key vector. Represents the enhanced value vector;

[0083] In the process of obtaining the output features of convolution-attention feature fusion through 1×1×1 convolution based on the outputs of the global attention branch and the local convolution branch, the following relationship exists:

[0084] ;

[0085] in, This represents the output feature of the convolution-attention feature fusion. This indicates that the process has undergone a 1×1×1 convolution.

[0086] After concatenating all the output features processed by the cascaded attention mechanism along the channel dimension and then performing linear projection to obtain the final output features of the cascaded attention mechanism, the following relationship exists:

[0087] ;

[0088] in, This represents the final output feature of the cascaded attention mechanism. This indicates that the process has undergone linear projection. This represents the output characteristics of each group after processing by the cascaded group attention mechanism. Indicates the total number of groups;

[0089] In the process of obtaining the enhanced multi-scale feature map by utilizing the final output features of the cascaded attention mechanism and the given feature map through linear transformation and 1×1 convolution, the following relationship exists:

[0090] ;

[0091] in, This represents the enhanced multi-scale feature map. Represents a given feature map, This indicates that the process has undergone a linear transformation.

[0092] Furthermore, this process enables recursive fusion of multi-head contexts, and the introduction of a convolutional-attention feature fusion mechanism significantly enhances the local-global semantic modeling capability.

[0093] Step 5: After the enhanced multi-scale feature map passes through the neck and head, the target recognition result is obtained;

[0094] In step 5, the enhanced multi-scale feature map passes through the neck and head to obtain the target recognition result. In this process, the target detection network uses an adaptive shared convolutional head to replace the original detection head structure.

[0095] Furthermore, the neck is a multi-scale feature fusion network located between the backbone network and the detection head. Through a bidirectional feature transfer path from top to bottom and bottom to top, it performs cross-scale fusion processing on feature maps from different levels of the backbone network to enhance the network's ability to detect targets of different sizes.

[0096] The head is the output module of the object detection network, used for object category prediction and bounding box regression based on the fused multi-scale feature map; in this invention, the object detection network uses an adaptive shared convolutional head (ASC-Head) to replace the original detection head structure.

[0097] The original detection head structure was replaced with an ASC-Head, which introduces a shared convolutional module to extract general feature representations. This design significantly reduces the number of parameters. In this way, the ASC-Head significantly reduces computational complexity while maintaining recognition accuracy, and ultimately outputs the target's category and location information.

[0098] For further details, please refer to Figure 6 , Figure 6 In the diagram, P3, P4, and P5 represent three feature maps of different scales from the neck network. P3 corresponds to the high-resolution feature for small target detection, P4 corresponds to the medium-resolution feature for medium-sized targets, and P5 corresponds to the low-resolution feature for large targets.

[0099] Conv_CN 1X1 represents a 1×1 convolutional layer with group normalization, used to project input features of different scales onto a unified feature space; Conv_CN 3X3 represents a 3×3 shared convolutional layer with group normalization, which is the core component of ASC-Head. It processes features of all scales through a parameter sharing mechanism, significantly reducing the number of model parameters.

[0100] Conv_Reg represents the regression branch convolutional layer, used to predict the bounding box location of the target; Conv_Cls represents the classification branch convolutional layer, used to predict the class probability of the target; Scale represents the adaptive scaling factor, which is a learnable parameter (initialized to 1.0) learned independently for each scale, used to dynamically adjust the contribution weight of prediction results at different scales to compensate for scale representation differences that may be caused by parameter sharing.

[0101] This invention provides some test data for verification. Experiments were conducted on a self-built ASD dataset and the publicly available M3FD dataset. The experimental results, compared with other state-of-the-art methods, show that SF-YOLO performs best on the self-built ASD dataset, achieving an mAP50 of 89.6%, as shown in Table 1.

[0102] Table 1. Comparison of results of different algorithms on the ASD dataset

[0103]

[0104] The present invention achieves an mAP50 of 82.3% on the publicly available multimodal target recognition dataset M3FD, reaching an excellent level and verifying the generalization performance of the proposed algorithm, as shown in Table 2:

[0105] Table 2. Comparison of results of different algorithms on the M3FD dataset

[0106]

[0107] Please see Figure 7 , Figure 7 This section compares the recognition results of different algorithms on cross-sectional images of complex indoor scenes. The top shows two sets of 3D models and the positions of their cutting planes. Figure 7 The circled numbers 1, 2, 3, and 4 represent the cutting plane indices at four different locations, and the corresponding cross-sectional views below are extracted from the 3D model.

[0108] The second line (Ground Truth): Human-annotated real labels, used as the evaluation benchmark;

[0109] The third row (YOLO11): The recognition results of the single-input baseline model YOLO11;

[0110] The fourth row (CFT) shows the recognition results of the multimodal fusion algorithm CFT; the fifth row (Ours) shows the recognition results of the SF-YOLO algorithm proposed in this invention. Red triangles in the figure indicate recognition errors, including missed detections (no target detected) and false positives (misclassification).

[0111] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0112] In the description of this specification, the terms "one embodiment," "some embodiments," and "Example 7" are used to refer to specific features, structures, materials, or characteristics that are included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0113] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A profile target recognition method based on cross-attention guided feature fusion, characterized in that, The method includes the following steps: Step 1: Select 3D models from the dataset and randomly combine them into an indoor scene model using modeling software; cut the indoor scene model along the same axis to obtain the ASD profile dataset. Step 2: Input the ASD profile dataset into the dual-input backbone network to extract features from two consecutive profiles in parallel, resulting in a dual-branch extracted feature map. Step 3: Perform cross-attention guided feature fusion on the feature maps extracted from the two branches to fuse the complementary information of the two profile maps, resulting in a fused feature map. The specific steps are as follows: The feature map extracted from the first branch is aligned by channel dimension using an adaptive transformation function to obtain a profile feature map after adaptive feature alignment. After the profile feature map after adaptive feature alignment and the feature map extracted by the second branch are concatenated along the channel dimension, an adaptive weighting process is then applied to obtain the feature map weighted by the attention mechanism. By using the feature maps weighted by the attention mechanism and the feature maps extracted from each branch, and through the cross-branch information interaction mechanism, we obtain the cross-sectional view feature maps of the two branches after cross-branch residual enhancement. Using the cross-branch residual enhanced cross-sectional view feature maps of the two branches, interactive features are obtained through element-wise multiplication. The interaction features are processed sequentially through convolutional projection, batch normalization, and sigmoid activation to obtain context-guided weights. By using context-guided weights, the cross-sectional view feature map after cross-branch residual enhancement is stitched together along the channel dimension to obtain the fused feature map. Step 4: Perform multi-scale feature enhancement processing on the fused feature map to obtain the enhanced multi-scale feature map; Step 4 combines the cascaded attention mechanism and the convolution-attention feature fusion mechanism. The specific steps are as follows: The fused feature maps are divided into groups, and 1×1 convolutions are used to jointly extract the feature maps in each group; Based on the output of the global attention branch and the output of the local convolution branch, through Convolution processing yields output features fused with convolutional and attention features. The output features of the convolution-attention feature fusion are used as group outputs. The current group output is added to the feature map in the next group element by element to obtain the output features processed by the cascaded group attention mechanism. After concatenating all the output features processed by the cascaded attention mechanism along the channel dimension, a linear projection is performed to obtain the final output features of the cascaded attention mechanism. Utilizing the final output features of the cascaded attention mechanism and a given feature map, through linear transformation and... After convolution processing, an enhanced multi-scale feature map is obtained; Step 5: After the enhanced multi-scale feature map passes through the neck and head, the target recognition result is obtained.

2. The profile target recognition method based on cross-attention guided feature fusion according to claim 1, characterized in that, In step 2, the ASD profile dataset is input into a dual-input backbone network to extract features from two consecutive profiles in parallel, resulting in a dual-branch extracted feature map. The corresponding process has the following relationship: ； in, This represents the feature map extracted by the first branch. This represents the feature map extracted by the second branch. This indicates that the feature extraction process is performed through the first branch. This indicates that feature extraction is performed through the second branch. This represents the input to the first branch. This represents the input for the second branch.

3. The profile target recognition method based on cross-attention guided feature fusion according to claim 2, characterized in that, In the process of using an adaptive transformation function to align the channel dimensions of the feature map extracted from the first branch, resulting in the adaptively aligned profile feature map, the following relationship exists: ； in, This represents the profile feature map after adaptive feature alignment. This indicates that the process is handled by an adaptive transformation function.

4. The profile target recognition method based on cross-attention guided feature fusion according to claim 3, characterized in that, In the process of obtaining interactive features by using the cross-branch residual-enhanced section view feature map and performing element-wise multiplication, the following relationship exists: ； in, Indicates interactive features, This represents element-wise multiplication. This represents the first cross-sectional view feature map after cross-branch residual enhancement. This represents the feature map of the second cross-sectional view after cross-branch residual enhancement; In the process of processing the interaction features sequentially through convolutional projection, batch normalization, and sigmoid activation to obtain the context-guided weights, the following relationship exists: ； in, Indicates context-guided weights, This indicates that the signal has been processed using the sigmoid activation function. This indicates that the batch normalization process has been performed. This indicates that the image has undergone 1×1 convolutional projection processing. In the process of using context-guided weights to perform channel-dimensional stitching on the cross-branch residual enhancement feature map to obtain the fused feature map, the following relationship exists: ； in, This represents the fused feature map.

5. The profile target recognition method based on cross-attention guided feature fusion according to claim 4, characterized in that, In step 4, the fused feature map is subjected to multi-scale feature enhancement processing to obtain an enhanced multi-scale feature map, which specifically includes the following steps: The fused feature maps are divided into groups, and 1×1 convolutions are used to jointly extract the feature maps in each group to obtain query vectors, key vectors, and value vectors. Attention is calculated using query vectors, key vectors, and value vectors to obtain the output of the global attention branch; The query vector, key vector, and value vector are subjected to depthwise separable convolution to obtain the enhanced query vector, enhanced key vector, and enhanced value vector, respectively. By integrating the enhanced query vector, key vector, and value vector through a fully connected layer, the output of the local convolutional branch is obtained.

6. The profile target recognition method based on cross-attention guided feature fusion according to claim 5, characterized in that, In the process of jointly extracting the feature maps from each group using 1×1 convolution to obtain the query vector, key vector, and value vector, the following relationship exists: ； in, Represents the query vector. Represents the key vector. Represents a value vector. Indicates passage Convolutional joint extraction processing, Indicates the first Feature maps in each group; In the process of calculating attention using query vectors, key vectors, and value vectors to obtain the output of the global attention branch, the following relationship exists: ； in, This represents the output of the global attention branch. This represents the transpose of the key vector. This represents the temperature parameter of the science department. This represents the number of channels for each self-attention head; In the process of integrating information through a fully connected layer using the enhanced query vector, enhanced key vector, and enhanced value vector to obtain the output of the local convolutional branch, the following relationship exists: ； in, This represents the output of a local convolution branch. This indicates that the data has undergone a 3×3×3 depthwise convolution. This indicates that the process has been handled by a fully connected layer. This indicates a feature concatenation operation. This represents the enhanced query vector. This represents the enhanced key vector. This represents the enhanced value vector.

7. The profile target recognition method based on cross-attention guided feature fusion according to claim 6, characterized in that, In the process of obtaining the output features of convolution-attention feature fusion through 1×1×1 convolution based on the outputs of the global attention branch and the local convolution branch, the following relationship exists: ； in, This represents the output feature of the convolution-attention feature fusion. This indicates that the process has undergone a 1×1×1 convolution. After concatenating all the output features processed by the cascaded attention mechanism along the channel dimension and then performing linear projection to obtain the final output features of the cascaded attention mechanism, the following relationship exists: ； in, This represents the final output feature of the cascaded attention mechanism. This indicates that the process has undergone linear projection. This represents the output characteristics of each group after processing by the cascaded group attention mechanism. Indicates the total number of groups; In the process of obtaining the enhanced multi-scale feature map by utilizing the final output features of the cascaded attention mechanism and the given feature map through linear transformation and 1×1 convolution, the following relationship exists: ； in, This represents the enhanced multi-scale feature map. Represents a given feature map, This indicates that the process has undergone a linear transformation.

8. The profile target recognition method based on cross-attention guided feature fusion according to claim 7, characterized in that, In step 5, the enhanced multi-scale feature map passes through the neck and head to obtain the target recognition result. In this process, the target detection network uses an adaptive shared convolutional head to replace the original detection head structure.

Images