A cross-modal pedestrian re-identification model, product and device
By using a hybrid model based on Transformer and ResNet, the problem of data heterogeneity in cross-modal pedestrian re-identification is solved, achieving efficient pedestrian identification in complex environments, improving the identification accuracy of public security work and the application value of smart cities.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINESE PEOPLE'S PUBLIC SECURITY UNIVERSITY
- Filing Date
- 2025-07-26
- Publication Date
- 2026-07-07
AI Technical Summary
Traditional visible light cameras are difficult to work effectively at night, in foggy or low-light environments, while infrared, thermal imaging and other modal devices have significant cross-modal data heterogeneity due to differences in imaging principles. Existing cross-modal pedestrian re-identification technologies lack matching accuracy in all-weather monitoring.
We employ a hybrid model based on Transformer and ResNet, combining local feature extraction and global semantic modeling capabilities. By using residual connections and attention mechanisms to enhance feature representation, we achieve the fusion and interaction of multimodal features and output high-level features for task prediction.
It improves the accuracy of cross-modal pedestrian matching in complex environments, adapts to harsh conditions such as nighttime, enhances the identification capabilities in public security work, reduces labor costs, and has application value in smart city construction.
Smart Images
Figure CN120954112B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a cross-modal pedestrian re-identification model, product, and device based on Transformer and ResNet. Background Technology
[0002] With the rapid development of intelligent security and smart city construction, the demand for all-weather, multi-scenario pedestrian tracking in the public safety field is becoming increasingly urgent. Traditional visible light cameras are difficult to operate effectively at night, in foggy or low-light environments, while infrared and thermal imaging devices, although able to overcome lighting limitations, suffer from significant cross-modal data heterogeneity due to differences in imaging principles. For example, visible light images rely on color and texture information, while infrared images primarily reflect the temperature distribution of the target. Cross-Modality Person Re-ID (CPR) technology aims to achieve accurate matching of the same pedestrian in different modalities through collaborative analysis of heterogeneous modal data, becoming a key technology for solving the challenge of all-weather monitoring. As public safety needs become increasingly complex, CPR technology demonstrates irreplaceable value in public security work, and its research and optimization have become one of the important topics in this field. Summary of the Invention
[0003] The present invention aims to at least partially solve one of the technical problems in the aforementioned related technologies.
[0004] Therefore, the purpose of this invention is to provide a cross-modal pedestrian re-identification model, product, and device based on Transformer and ResNet, which has both local feature extraction and global semantic modeling capabilities, can improve the accuracy of cross-modal pedestrian matching, and can adapt to complex environments such as nighttime.
[0005] To solve the above-mentioned technical problems, the present invention is implemented as follows:
[0006] This invention provides a cross-modal person re-identification model, which is a hybrid model based on ResNet50 and Transformer; the model includes:
[0007] The image input module is configured to provide basic data and perform preliminary extraction and interaction of cross-modal features;
[0008] The feature processing module is configured to perform splitting, convolution, and integration of features after multimodal fusion, thereby achieving preliminary feature processing.
[0009] The residual and attention module is configured to solve network training problems through residual connections, while combining the attention mechanism to select key features and improve feature representation capabilities.
[0010] The feature enhancement and fusion module is configured to enhance and integrate deep features, fusing multi-source information to output more comprehensive high-level features;
[0011] The output and post-processing module is configured to normalize and optimize the fused features, and output the results for task prediction.
[0012] In addition, the cross-modal pedestrian re-identification model according to the present invention may also have the following additional technical features:
[0013] In some embodiments, the image input module includes:
[0014] The VIS branch and IR branch are configured to perform preliminary processing on the input data of the corresponding modality, guide the data flow and mine early features;
[0015] Two MIE modules are configured to extract basic features from single-modal VIS and IR data, respectively, to prepare for cross-modal interaction.
[0016] MID module: Receives features from VIS and IR branches after MIE processing, realizes the complementarity and enhancement of features from different modalities, outputs preliminary features after fusion, and alleviates the problem of cross-modal data heterogeneity.
[0017] In some embodiments, the feature processing module includes:
[0018] The Vtm branch is configured to extract local features from VIS-related data through convolutional layers.
[0019] The Itm branch is configured to extract local features from IR-related data using convolutional layers.
[0020] In some implementations, the residual and attention module includes several alternately arranged residual blocks and coordinate attention modules:
[0021] Residual blocks are configured to solve the gradient vanishing and degradation problems in deep networks through residual connections, preserve shallow feature information and learn complex features, and ensure the stability of feature propagation.
[0022] The coordinate attention module is configured to extract spatial coordinate information of features, explore dependencies between channels, accurately capture feature details, and enable the network to focus on valuable feature regions.
[0023] In some of these implementations, the feature enhancement and fusion module includes:
[0024] The MPGMCA module is configured to enhance the attention mechanism of the channel dimension, refine the feature weight allocation of different modalities and channels, and highlight key channels.
[0025] The Transformer module is configured to mine long-distance feature dependencies through a self-attention mechanism, model global semantics, and improve the global correlation of features.
[0026] HFF module: It is configured to aggregate the output features of residual and attention modules and the output features of the Transformer module, and fuse information from different levels and mechanisms to output comprehensive high-level features.
[0027] In some embodiments, the output and post-processing module includes:
[0028] AP module: It is configured to adjust feature dimensions, compress spatial size, retain key features, and prepare for subsequent processing;
[0029] BN: It is configured to normalize features, accelerate network convergence, improve training stability, and optimize feature distribution;
[0030] Multiple task output heads: configured to receive features after BN processing, complete the final prediction for a specific task, and achieve the goal of cross-modal person re-identification.
[0031] In some of these implementations, the HFF module includes:
[0032] The Gi and Li modules are configured to accept corresponding feature data as input.
[0033] The core processing branch is configured to include parallel max pooling and average pooling layers to pool the Gi input data, and then merge them and pass them through a shared multilayer perceptron, a sigmoid activation function and channel attention processing to obtain channel attention weights. Then, the key channel features are enhanced by feature weighting of the original Gi channel input.
[0034] The core processing branch is configured to include a max pooling layer and an average pooling layer connected in series to pool the Li input data. After processing by a convolutional module and a sigmoid activation function, spatial attention weights are obtained. Then, the original input of the Li channel is weighted to enhance key spatial location features.
[0035] The fusion output section is configured to perform residual-connected enhanced gradient processing on the spliced upper and lower core processing branch data, and then adjust the dimensions to output regular features.
[0036] In some implementations, the coordinate attention module includes:
[0037] Coordinate information embedding module: It is configured to capture global features in the column and row directions respectively through average pooling in the X and Y directions, and incorporate them into the features;
[0038] Coordinate attention generation module: It is configured to concatenate, convolve, batch normalize and non-linearly activate the pooling results to generate attention weights in the X and Y directions;
[0039] Feature weighting: It is configured to multiply attention weights with the original input features element by element to achieve feature weighting, thereby strengthening key spatial location features, suppressing irrelevant regions, and outputting a feature map that focuses on effective information.
[0040] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the content of the cross-modal pedestrian re-identification model as described in any of the preceding embodiments.
[0041] This invention also provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the content of the cross-modal pedestrian re-identification model as described in any of the preceding embodiments.
[0042] Compared with the prior art, the present invention has at least the following beneficial effects:
[0043] In this embodiment of the invention, the provided cross-modal person re-identification model based on Transformer and ResNet, with ResNet50 network as the baseline model, fully extracts local features. In the experiment, the Transformer-based model is improved and global features are tested. This model integrates the two, enabling the fusion network to extract both local and global features. It also incorporates a coordinate attention module, a multi-feature generation module, and an HFF module to further improve the accuracy of the model.
[0044] In this embodiment of the invention, the cross-modal person re-identification model based on Transformer and ResNet is an improvement on the ResNet50 and Transformer model. The new model combines the inherent advantages of ResNet50 in capturing local features with the unique ability of Transformer in modeling global context. It fully leverages the synergistic effect of the hybrid architecture, using ResNet50 to extract local details (such as clothing texture) in a shallow layer and Transformer to model global semantics (such as the overall outline of the pedestrian) in a high-level layer. CNN is responsible for spatial dimension features, and Transformer performs attention reweighting in the channel dimension.
[0045] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0046] Figure 1 This is a schematic diagram of the overall framework of a model disclosed in one embodiment of the present invention;
[0047] Figure 2 This is a structural diagram of an HFF module disclosed in one embodiment of the present invention;
[0048] Figure 3 This is a structural diagram of a multi-feature generation module disclosed in one embodiment of the present invention;
[0049] Figure 4 This is a coordinate attention structure diagram disclosed in one embodiment of the present invention. Detailed Implementation
[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0051] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific examples and application scenarios.
[0052] In some embodiments of this invention, a cross-modal pedestrian re-identification model based on Transformer and ResNet is provided. This hybrid model, based on ResNet50 and Transformer, combines local feature extraction with global semantic association, offering an innovative breakthrough in solving technical challenges in public security operations. Its significance for public security work lies primarily in its highly efficient identification capabilities in nighttime / adverse environments: traditional visible light cameras experience a significant drop in recognition rates in low light, fog, or nighttime scenes. By extracting local thermal radiation features (such as head contours and gait) from infrared images using ResNet50, and combining this with Transformer modeling the global correlation between visible light and infrared modalities, pedestrian identification can be achieved even in complete darkness. This facilitates case review and suspect identification by public security organs, key area deployment and early warning in security operations, reduces manpower costs, and improves law enforcement accuracy. It also benefits the construction of smart cities and nationwide surveillance projects, possessing immense value for social development.
[0053] In some embodiments of this invention, a hybrid model based on ResNet50 and Transformer is used; please refer to [link / reference needed]. Figure 1 As shown, it includes an image input module, a feature processing module, a residual and attention module, a feature enhancement and fusion module, and an output and post-processing module.
[0054] The image input module is the data source for the entire process, providing raw data in visible and infrared modes to lay the foundation for subsequent multimodal fusion processing. The image input module includes:
[0055] The VIS branch and IR branch are channels for preliminary processing of the input data of the corresponding modes, and each is responsible for the preliminary feature mining and flow guidance of the visible light mode and infrared mode data.
[0056] Two MIE modules (Multi-Modal Information Extraction), corresponding to the VIS branch and the IR branch respectively, are used to extract basic features from single-modal (visible light or infrared) data, prepare for subsequent cross-modal interaction, and mine valuable information in single-modal data;
[0057] MID (Multi-Modal Interaction and Distillation): Receives features from VIS and IR branches after MIE processing, realizes cross-modal feature interaction, allows different modal features to complement each other, and may also use a "distillation" mechanism to filter and strengthen effective features, outputting preliminary features after multimodal fusion.
[0058] The feature processing module is used to further split, convolve, initially integrate, and transform features from multiple modalities. The feature processing module includes:
[0059] The Vtm branch and its corresponding convolutional layer are used to perform convolution operations on the combination of VIS branch data and preliminary feature data after multimodal fusion. Local features are extracted through convolution kernels, feature dimensions are changed, the number of channels is compressed / expanded, and features are initially integrated and transformed.
[0060] The Itm branch and its corresponding convolutional layer are used to perform convolution operations on the combination of IR branch data and preliminary feature data after multimodal fusion. Local features are extracted through convolution kernels, feature dimensions are changed, and the number of channels is compressed / expanded, thus initially integrating and transforming features.
[0061] The residual and attention module includes several alternately arranged residual blocks and coordinate attention modules. One residual block and one coordinate attention module constitute a group, preferably four groups, which are connected in sequence. That is, the output of the previous group is used as the input of the next group; within the same group, the output of the residual block is used as the input of the coordinate attention module.
[0062] Residual blocks address the vanishing and degradation issues during deep network training, enabling the network to learn more complex features while preserving shallow feature information, thus ensuring the stability and effectiveness of feature transfer.
[0063] The coordinate attention module, working in conjunction with the residual block, focuses on the spatial coordinate information of features when processing features, uncovers dependencies between channels, accurately captures feature details, improves the ability to extract key features, and makes the network pay more attention to valuable feature regions.
[0064] Feature enhancement and fusion modules, including:
[0065] The MPGMCA module (Multi-Path Guided Multi-Modal Channel Attention) further strengthens the attention mechanism of multimodal features in the channel dimension. Combined with multi-path guidance, it refines the interaction and weight allocation of features of different modalities and channels, highlighting key channel features.
[0066] Transformer module: Utilizes self-attention mechanism to mine long-distance feature dependencies, models global features, integrates feature information scattered in different locations, and improves the global correlation of features;
[0067] The HFF module (High-level Feature Fusion Block) fuses multiple features processed by the preceding modules (features from residual blocks, attention, MPGMCA, and Transformer), summarizing and fusing features extracted from different levels and mechanisms to output more comprehensive high-level features.
[0068] The MPGMCA module and the Transformer module are set up in sequence to process the output of the second-to-last set of residual blocks and the coordinate attention module, and input the data obtained together with the output of the last set of residual blocks and the coordinate attention module into HFFBlock.
[0069] Output and post-processing modules, including:
[0070] Output (Feature Output): The final feature representation after HFFBlock fusion, providing a high-dimensional feature foundation that integrates multimodal information for subsequent tasks (such as recognition, classification, etc.);
[0071] AP (Adaptive Pooling): Performs pooling operations on the output features to adjust the feature dimensions, compress the spatial size, and retain key features, preparing for subsequent operations such as BN, making the features more regular.
[0072] BN (Batch Normalization): Normalizes the features after AP processing, which accelerates network convergence, improves training stability, reduces internal covariate bias, and makes the feature distribution more reasonable.
[0073] Ltri, Lid, and Ldci are different task output heads (such as different category recognition, different modality-specific tasks, etc.). They receive the features after BN processing and perform the final prediction output for a specific task (such as classification, recognition, and specific modality-related tasks) to complete the task objective of the entire multimodal processing flow.
[0074] The multimodal (visible light + infrared) feature fusion and processing network achieves a complete process from single-modal feature extraction and cross-modal interaction to deep feature enhancement and fusion, and finally outputs a result for task prediction through multi-branch and multi-module collaboration.
[0075] In some embodiments of the present invention, the structure of the HFF module is as follows: Figure 2 As shown, it includes the input section (Gi, Li), two core processing branches (upper and lower), and the fused output section. Among them:
[0076] Gi is the primary feature input (basic feature map from the backbone network), containing rich spatial and channel information. Li is the auxiliary information input (low-level features, label guidance information, or other modal features), used to supplement / guide feature processing.
[0077] The aforementioned core processing branch is a channel attention + feature weighting branch, including a max pooling layer and an average pooling layer (Maxpool and Avgpool in the diagram). These two pooling layers are set up side-by-side to pool the Gi input data separately, and then the data is merged and fed into a Shared Multilayer Perceptron (Share MLP). The Shared MLP performs "compression-expansion" (dimensionality reduction followed by dimensionality increase) on the pooling results, learning the non-linear dependencies between channels and identifying the importance of key channels. The data then passes through a Sigmoid activation function and a channel attention mechanism. The Sigmoid function maps the MLP output to [0,1], yielding channel attention weights to measure the contribution of each channel to the task. Multiplying the output channel attention weights by the original input of the Gi channels achieves feature weighting, strengthening the features of key channels and suppressing unimportant channels.
[0078] The aforementioned core processing branch is a spatial attention + feature selection branch, including max pooling layers and average pooling layers, a 7×7 convolutional module, a sigmoid activation function, and a feature weighting module. Max pooling and average pooling layers are set sequentially; the processed data is then processed by the convolutional module to extract global spatial context using large convolutional kernels. It is then processed by the sigmoid activation function to output spatial attention weights. These spatial attention weights are multiplied by the original input (Li) to achieve feature weighting, with element-wise multiplication to enhance key spatial location features.
[0079] The outputs of the two core processing branches are concatenated and then input into the convolutional residual module. This module first performs a 3×3 convolution to extract local features, and then connects them with the concatenated data (⊕) to preserve the original information and avoid gradient vanishing. Finally, the data undergoes several 1×1 convolutions to adjust the dimensions and output refined features. Specifically, the dimensions can be adjusted twice: the first 1×1 dimensionality increase (dim×4) enhances the expression, and the second 1×1 dimensionality reduction (1×dim) restores the number of channels. The final output of the regularized features is fed into the multi-feature generation module Fi.
[0080] The upper branch uses channel attention weighting to highlight key channels, while the lower branch uses spatial attention weighting to highlight key regions. This dual-dimensional filtering method can obtain key features, while residual and MLP are used to ensure the effectiveness of feature learning. This allows the model to more accurately capture information valuable to the task and improve the performance of tasks such as classification and detection.
[0081] In some embodiments of the present invention, the multi-feature generation module Fi is used to generate diverse and refined features from the input features, and its structure is as follows: Figure 3 As shown, the model includes several feature generation branches, each comprising a cross-convolution module, a coordinate attention module, and a fully connected layer arranged sequentially. The cross-convolution module consists of three parallel cross-convolutions with different kernels to extract features at multiple scales and from multiple perspectives. The outputs of the three cross-convolutions are summed to supplement information at different scales and avoid feature fragmentation. The summed data is then input into the coordinate attention module, which includes parallel channel attention and spatial attention. Channel attention assigns higher weights to key channels and suppresses irrelevant channels; spatial attention highlights the target region, weakens the background, and accurately captures spatial details. Together, they refine features from both channel and spatial dimensions, allowing the model to focus more on valuable information. The features processed by channel and spatial attention are then concatenated to aggregate dual-attention information and retain a more comprehensive feature representation. The concatenated features are then input into the fully connected layer for dimensionality transformation (channel compression / expansion) to further integrate global information and output well-structured features. The multi-feature generation module is a feature enhancement module that combines "multi-branch mining + dual attention refinement + full connection integration". It can make the output features richer and more focused on key information to serve downstream tasks.
[0082] In some embodiments of the present invention, the structure of the coordinate attention module, which is alternately arranged with the residual block in the residual and attention module, is as follows: Figure 4 As shown, the system includes a coordinate information embedding module and a coordinate attention generation module. X is the input feature map, including spatial and channel information. The coordinate information embedding module includes an X-axis average pooling layer and a Y-axis average pooling layer, set side-by-side. For the input feature map X, features are compressed along the X-axis, retaining vertical (Y-axis) dimensional information and capturing global features in the column direction; features are compressed along the Y-axis, retaining horizontal (X-axis) dimensional information and capturing global features in the row direction. Coordinate (position) information is embedded through average pooling in both directions, allowing the model to perceive the distribution pattern of features along the X / Y axes. The column and row data are input into the coordinate attention generation module, first undergoing concatenation and convolution operations, then batch normalization and non-linear activation, and then split into two paths, each undergoing convolution and activation to obtain two sets of attention weights. These two sets of attention weights are multiplied by the original input feature map X to achieve weighted summation, and then fused to output a weighted feature map Y. When splitting into two paths, the attention weights for rows and columns are obtained at the concatenation point.
[0083] In some embodiments of the present invention, experiments were conducted using the method of the present invention, including: experiments were conducted on a server equipped with an NVIDIA Tesla A100 GPU, using the PyTorch deep learning framework; when training the MMN model, the resolution of the pedestrian images was first uniformly adjusted to 384 pixels wide and 192 pixels high, and the number of epochs was set to 200; the batch size was set to 256, including 4 different types of pedestrian images, with 8 images for each pedestrian. In the cross-modal hard samples, the 8 images for each pedestrian included 4 visible light images and 4 infrared images; the initial learning rate was set to 0.01, and then it was linearly increased to 0.1 over 10 epochs. Then, the learning rate was decayed to 0.001 over 20 epochs, and then further decayed to 0.0001 in each small batch. This was optimized using the Adam optimizer with a momentum parameter set to 0.9.
[0084] In some embodiments of the present invention, after inputting images in both infrared and visible light modes, infrared branches, visible light branches, and a third branch (m-modality) are generated together with the original VIS and IR images to produce an m-modality image. This m-modality image is then fed into a two-stream ResNet50 backbone to extract modality-invariant features. The first convolutional block in each stream is different to capture modality-specific low-level representations, while intermediate and deep convolutional blocks are shared. A new branch, MFGFGMCA, is introduced after the third residual network and then fed into the Transformer. Both streams enter the HFF module. The final output model divides the feature map horizontally into several parts and feeds each part into a classifier to learn local cues. The stride of the last convolutional block is further modified to 1 inch to preserve more spatial information in the backbone. After the convolutional layers with average pooling (AP) layers, a batch normalization (BN) layer with shared parameters across all modality images is added to facilitate loss convergence. Finally, the features before and after the BN layer are fed into different loss functions to collectively improve the network.
[0085] All parts of this invention not described in detail herein can be referred to in the prior art or are known to those skilled in the art. This embodiment does not limit these aspects and will not describe them in detail here.
[0086] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these modifications are within the protection scope of the present invention.
Claims
1. A cross-modal pedestrian re-identification model, characterized in that, The model is a hybrid model based on ResNet50 and Transformer; the model includes: The image input module is configured to provide basic data and perform preliminary extraction and interaction of cross-modal features; The feature processing module is configured to perform splitting, convolution, and integration of features after multimodal fusion, thereby achieving preliminary feature processing. The residual and attention module is configured to solve network training problems through residual connections, while combining the attention mechanism to select key features and improve feature representation capabilities. The feature enhancement and fusion module is configured to enhance and integrate deep features, fusing multi-source information to output more comprehensive high-level features; The output and post-processing module is configured to normalize and optimize the fused features and output results for task prediction. The feature enhancement and fusion module includes: The MPGMCA module is configured to enhance the attention mechanism of the channel dimension, refine the feature weight allocation of different modalities and channels, and highlight key channels. The Transformer module is configured to mine long-distance feature dependencies through a self-attention mechanism, model global semantics, and improve the global correlation of features. HFF module: It is configured to aggregate the output features of residual and attention modules and the output features of the Transformer module, and fuse information from different levels and mechanisms to output comprehensive high-level features; The HFF module includes: The Gi and Li modules are configured to accept corresponding feature data as input. The core processing branch is configured to include parallel max pooling and average pooling layers to pool the Gi input data, and then merge them and pass them through a shared multilayer perceptron, a sigmoid activation function and channel attention processing to obtain channel attention weights. Then, the key channel features are enhanced by feature weighting of the original Gi channel input. The core processing branch is configured to include a max pooling layer and an average pooling layer connected in series to pool the Li input data. After processing by a convolutional module and a sigmoid activation function, spatial attention weights are obtained. Then, the original input of the Li channel is weighted to enhance key spatial location features. The fusion output section is configured to perform residual connection-based enhancement gradient processing on the spliced upper and lower core processing branch data, and then after dimension adjustment, output regular features. The image input module includes: The VIS branch and IR branch are configured to perform preliminary processing on the input data of the corresponding modality, guide the data flow and mine early features; Two MIE modules are configured to extract basic features from single-modal VIS and IR data, respectively, to prepare for cross-modal interaction. MID module: Receives features from VIS and IR branches after MIE processing, realizes the complementarity and enhancement of features from different modes, outputs preliminary features after fusion, and alleviates the problem of cross-modal data heterogeneity. The residual and attention module includes several alternately arranged residual blocks and coordinate attention modules: Residual blocks are configured to solve the gradient vanishing and degradation problems in deep networks through residual connections, preserve shallow feature information and learn complex features, and ensure the stability of feature propagation. The coordinate attention module is configured to extract spatial coordinate information of features, explore inter-channel dependencies, accurately capture feature details, and enable the network to focus on valuable feature regions. The coordinate attention module includes: Coordinate information embedding module: It is configured to capture global features in the column and row directions respectively through average pooling in the X and Y directions, and incorporate them into the features; Coordinate attention generation module: It is configured to concatenate, convolve, batch normalize and non-linearly activate the pooling results to generate attention weights in the X and Y directions; Feature weighting: It is configured to multiply attention weights with the original input features element by element to achieve feature weighting, thereby strengthening key spatial location features, suppressing irrelevant regions, and outputting a feature map that focuses on effective information.
2. The cross-modal pedestrian re-identification model according to claim 1, characterized in that, The feature processing module includes: The Vtm branch is configured to extract local features from VIS-related data through convolutional layers. The Itm branch is configured to extract local features from IR-related data using convolutional layers.
3. The cross-modal pedestrian re-identification model according to claim 1, characterized in that, The output and post-processing module includes: AP module: It is configured to adjust feature dimensions, compress spatial size, retain key features, and prepare for subsequent processing; BN: It is configured to normalize features, accelerate network convergence, improve training stability, and optimize feature distribution; Multiple task output heads: configured to receive features after BN processing, complete the final prediction for a specific task, and achieve the goal of cross-modal person re-identification.
4. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the content of the cross-modal pedestrian re-identification model as described in any one of claims 1-3.
5. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the contents of the cross-modal pedestrian re-identification model according to any one of claims 1-3.