A Domain-Generalized Person Re-identification Method and System Based on Meta-Graph Awareness

Abstract

A domain-generalized person re-identification method and system based on meta-graph awareness includes: a ResNet50 network as the backbone network, a meta-global correlation awareness module, a meta-graph relationship sampling module, and a person matching module. This invention applies the Meta-Graph Aware (M-GRA) algorithm to domain-generalized person re-identification. In the meta-training domain, all block features are stacked using an affinity model to construct pairwise relationships. Then, a shallow convolutional model is used to learn this feature stacking relationship model, and a global knowledge graph is constructed. This classifies and weights structural information, suppressing noise while maintaining learning efficiency, and preventing model overfitting.

Description

Technical Field

[0001] This invention relates to a technology in the field of image processing, specifically a domain-generalized pedestrian re-identification method and system based on Meta-Graph Awareness (M-GRA). Background Technology

[0002] Person re-identification (ReID), also known as pedestrian re-identification, aims to retrieve pedestrians with the same identity across different cameras in non-overlapping multi-camera scenarios, based on the target information of the pedestrian to be identified. It is a key technology in important applications such as multi-camera tracking and search and evidence collection. ReID technology strives to reduce the limitations of fixed cameras and has broad application prospects in intelligent surveillance, target tracking, and human-computer interaction. However, current research on person re-identification faces many challenges. Among them, domain generalization (DG) is a critical problem that ReID algorithms urgently need to solve. The training process cannot perceive target domain information; that is, it cannot maximize neighborhood invariant features and minimize the differences between two domains, leading to significant biases in the test results of the trained model. Large inter-domain distances are a significant factor contributing to the differences in pedestrian image features across different datasets. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies in reducing inter-domain differences and improving generalization performance when the target domain is inaccessible, as well as the inability to solve the domain drift problem between the source and target domains. It proposes a domain-generalized person re-identification method and system based on meta-graph awareness (M-GRA). The method applies the Meta-Graph Aware (M-GRA) algorithm to domain-generalized person re-identification. In the meta-training domain, all block features are stacked using an affinity model to construct pairwise relationships. Then, a shallow convolutional model is used to learn this feature stacking relationship model, and a global knowledge graph is constructed to classify and weight structural information. This approach suppresses noise while maintaining learning efficiency and preventing model overfitting.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a domain-generalized person re-identification method based on meta-graph perception. It constructs a meta-learning framework and, during the meta-training stage, inputs source domain training set images into the meta-global relevance (GA) module of the meta-learning framework without any person identification annotations. Based on the obtained person features, a global knowledge graph is constructed, and the distances between these person features and cluster centers are calculated to build a person feature relationship graph between adjacent classes. This graph is then input into the meta-graph relation sampling (GRS) module to sample and learn discriminative features from adjacent classes. Finally, an optimizer is used to optimize the model, calculating the meta-training loss function to obtain a pre-trained model. Finally, a meta-testing... In the trial phase, the target domain dataset is input into the spatial perception and channel perception units to obtain the target domain data features and calculate the meta-test loss function. In the online phase, pedestrian images generated by the trained meta-learning framework are used to train a deep neural network for pedestrian re-identification in an unsupervised manner. The model trained in the source domain using the meta-learning framework is used to verify the generalization effect of the model in the target domain. That is, the generalization ability is reflected by the accuracy in the target domain. Four domain generalization pedestrian re-identification benchmark experiments are set up for comparison, and the Market1501 and DukeMTMC-ReID datasets are alternately used as the source and target domains for training and testing.

[0006] The source domain training set images used are the Market1501 and DukeMTMC-ReID datasets.

[0007] The target domain datasets used are Market1501, DukeMTMC-ReID, and MTMS17 datasets.

[0008] The global knowledge graph is constructed as follows: in the meta-training domain, all features extracted by the meta-global relevance perception module are processed in blocks, stacked to construct paired relations, and then the paired relations are learned through shallow convolutional model units to construct the global knowledge graph.

[0009] The aforementioned meta-training domain and meta-test domain refer to the following: In each iteration, the data in the source domain is randomly divided into N-1 meta-training domains and one meta-test domain. The model is trained in the meta-training domain and the model in the meta-test domain is copied to simulate the training and testing process of domain generalization.

[0010] The pedestrian feature relationship graph is constructed as follows: at the beginning of each epoch, a graph is constructed for all pedestrians, and the nearest neighboring class is always sampled in a mini-batch for discriminative learning; the distance from all pedestrian features to the cluster center in various data feature graphs extracted from the deep neural network of the input global correlation perception mechanism is calculated, and the similarity of pedestrian features between domains is used to construct the pedestrian feature relationship graph between adjacent pedestrians.

[0011] The aforementioned meta-training loss function includes: intra-domain scattering loss and triplet loss.

[0012] The aforementioned meta-test loss function includes: the triplet loss function.

[0013] Technical effect

[0014] This invention, within the meta-learning framework, inputs a global correlation perception mechanism and graph relation sampling into the ResNet50 network, significantly improving robustness and accuracy compared to existing technologies. Rank-1 and mAP increase by 5.25% and 3.73%, respectively, which are significantly better than existing representative algorithms. Attached Figure Description

[0015] Figure 1 This is a flowchart of the present invention;

[0016] Figure 2 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0017] like Figure 2 As shown in this embodiment, a domain-generalized pedestrian re-identification system based on meta-graph perception is included, comprising: a ResNet50 network as the backbone network, a meta-global correlation perception module, a meta-graph relationship sampling module, and a pedestrian matching module. Specifically: In the meta-training domain, a pre-trained ResNet50 network is used as the backbone network. To reduce inter-domain differences, the meta-global correlation perception module first passes through a channel perception unit and then a spatial perception unit, constructing an interaction model (paired relationships) by stacking block features using an affinity model, calculating relationship-aware nodes, constructing a global knowledge graph, and enhancing feature representation capabilities by weighting structural information and suppressing noise. At the beginning of each batch in the meta-training domain, the meta-graph relationship sampling module constructs relationship graphs for adjacent classes between domains and trains them together to shorten the distance between domains. The pedestrian matching module calculates matching accuracy and precision using cumulative matching characteristic (CMC) and average precision (mAP) based on the loss function results.

[0018] The meta-global relevance perception module includes: a shallow convolutional model unit, a channel perception unit, and a spatial perception unit. The shallow convolutional model unit is a 1x1 convolutional layer used to reduce the dimensionality of the 2048-dimensional feature map to 512 dimensions using a 1x1 convolutional kernel. The channel perception unit selects important channels based on the channel perception feature node values ​​generated in the channel domain, obtaining the results for important channels. The spatial perception unit selects important spatial regions and predicts the most relevant spatial locations based on the spatial perception feature node values ​​generated in the spatial domain, obtaining the results for regions of interest. Meta-graph relationship sampling is then performed based on the obtained important channels and important information regions.

[0019] The affinity model mentioned above is specifically: the affinity model m from node a to node b. a,b Defined as the dot product affinity in the embedding space, i.e., Where: i = 1, ..., N, the spatial locations of the feature map are labeled as 1, ..., N, and the N feature nodes are represented as m. a ∈R C f s Embed a function for the node itself, δ s and These are two embedding functions implemented through 1×1 spatial convolutional layers, batch normalization (BN), and ReLU activation, namely δ. s (m a =ReLU(W) δ m a ), in: s1 is a predefined positive integer that controls the dimensionality reduction rate. Similarly, the affinity m from node b to node a is obtained. b,a =f s (m b ,m a ), where: using (m a,b ,m b,a To describe m a and m b A two-way relationship between them.

[0020] The interaction model is as follows: For the a-th feature node, with node identities b = 1, 2, ..., N, the interaction model between it and all nodes, stacked in a fixed order, is as follows: in: The affinity matrix represents the interaction model (paired relationships) between nodes in the network.

[0021] like Figure 1 As shown in the figure, this embodiment relates to a domain-generalized person re-identification method based on meta-graph perception of the above system, which includes the following steps:

[0022] Step 1: Randomly divide the source domain training set images into a meta-training domain and a meta-test domain.

[0023] Step 2, Meta-training stage: Using a pre-trained ResNet50 network as the backbone network, the source domain dataset images are input into the meta-global correlation perception module to extract various pedestrian features, then the distance from the pedestrian features to the cluster center is calculated, a feature relationship graph between adjacent classes is constructed, and then input into the meta-graph relationship sampling module to sample and learn discriminative features for adjacent classes, and then optimize through an optimizer to obtain the pre-trained model.

[0024] This embodiment inputs dataset images into the meta-global correlation perception module to more accurately identify the relevant information between two domains, which is more efficient. On this basis, the meta-graph relationship sampling module combines the feature itself and the global information related to this feature to form all node information in pairs for stacked learning, which can better ensure the relevance of feature positions and make the model more efficient.

[0025] Step 3: Use convolutional dimensionality reduction. Specifically, a global max pooling layer is used to obtain feature vectors. Each feature vector is used to form a C-dimensional feature map. This device reduces the dimensionality of the 2048-dimensional feature map to 512-dimensional feature map by using a 1*1 convolutional kernel.

[0026] Step 4: Input the feature map obtained in Step 3 into the meta-global correlation sensing module containing channel sensing units and spatial sensing units to obtain channel sensing feature node values ​​and spatial sensing feature node values, specifically including:

[0027] 4.1) The feature space information after average pooling and max pooling operations is input into the shared network to generate the channel-aware feature node matrix.

[0028] The shared network includes a hidden layer and a multilayer perceptron (MLP).

[0029] The channel-sensing feature node mentioned above refers to: D a =[A(f s (V a )), G s (V a )]and Where: A(·) is the global average pooling operation along the channel dimension, f s G is the embedding function for the feature itself. s f is a global relation embedding function. s and G s Both are implemented using spatial 1×1 convolution, with normalization and ReLU activation as f. s (V a =ReLU(W) f V a ), G s (V a =ReLU(W) G V a ), S1 represents the control of the dimensionality reduction rate and S1∈N + .

[0030] 4.2) Perform matrix multiplication between the channel sensing node matrix and the original features to obtain the multiplication result;

[0031] 4.3) Perform element-wise summation on the multiplication result matrix and the original features to obtain the channel-aware feature node values, specifically: y a =Sigmoid(W2ReLU(W1D) a )), where: W1 and W2 are implemented by 1×1 convolution and normalization, and W1 transforms the channel size to 1.

[0032] 4.4) The spatial perception unit performs two pooling operations, average pooling and max pooling, and then concatenates the obtained features and performs a 1×1 convolution. This unit is used to learn a spatial map of size H×W, where H is the height of the spatial map and W is the width of the spatial map.

[0033] 4.5) The average pooling features and max pooling features are concatenated using a 1*1 convolutional layer to obtain the spatially aware feature node values, specifically: Where: A(·) is the global flat pooling operation along the channel dimension, f s G is the embedding function for the feature itself. s f is a global relation embedding function. s and G s Both are implemented using spatial 1×1 convolution, where l is the Sigmoid function.

[0034] In step 4, meta-global relevance perception refers to: weighting and summing the global relevance perception information and the original output. That is, in the global scope, the a-th feature node of the input image is combined and superimposed with all its related feature nodes to form the global feature information of interest. Specifically, this includes: global feature information. Where: the correlation between the a-th feature node and other feature b is: The a-th feature node is V a In addition, it also includes the global scope structural information of the feature itself and its related features, as well as local original information, and the two are not in the same feature domain.

[0035] Step 5: Superimpose the channel-sensing feature node values ​​and the spatial-sensing feature node values ​​in pairs to obtain the relation-sensing node K. Δ =F 1×1 (V a V b ) is used to represent global structure information, where: F 1×1 The operation is a 1×1 convolution; then a global information graph is learned and constructed through a shallow convolution model.

[0036] Step 6: Based on the global information graph and the pedestrian features obtained in Step 2, calculate the distance from all pedestrian features to the cluster center. Pedestrian images with similar distances are grouped together to construct a feature relationship graph between adjacent classes.

[0037] Step 7: Perform graph relation sampling (including positive sample cyclic relation sampling and negative sample random relation sampling) on ​​the feature relation graph between adjacent classes. Specifically, in each iteration, the source domain SD is randomly divided into the meta-training domain SD. tr Heyuan Test Domain SD te In the meta-training domain, the updated model is used to evaluate the distances between different clusters that provide the most clustering information. A graph is then constructed for the relevant clusters, and a random graph is selected to form a set Q. Finally, features V′∈Q are extracted. n×n Where: n is the total number of training classes, c is the feature dimension, and then the pairwise distance d∈Q of related classes is calculated using the cosine function. N×N For each class g, the constructed graph is: G = (I-1, E), where: I-1 = {g | g = 1, ..., g} are vertices, and each vertex can be searched for its neighboring class N(g), where N(g) = {V a ′|a=1,······I-1}, where: I is the number of classes to be sampled in each batch, and E is the edge, with the formula E={(g1,g2)g2∈N(g1)}.

[0038] For graph relation sampling, retrieve all related classes in the graph for each class, and then combine them with the class itself H to obtain a set Z = {H}∪{V′|(H,V)∈E}, where |Z| = I. Then, randomly sample each class in Z, with the mini-batch remaining unchanged for each epoch.

[0039] Step 8: Calculate the loss function: Use intra-domain scattering loss and triplet loss in the meta-training loss, and use triplet loss in the meta-testing loss to balance the overall parameter changes, finally obtaining the total meta-training loss and the total meta-testing loss. Update the overall parameters based on the total loss, starting from the meta-training domain SD. tr Sampling a small sample B r , from ε ρ Updated to In the meta-test domain SD te Unknown class sample B e Above, the overall parameter changes are balanced using triplet loss. Use this equation to perform a meta-update of the global parameters.

[0040] The total meta-training loss L o丈er =L S +L h +L tr Among them: intra-domain scattering loss Intra-domain scattering loss causes inter-domain and intra-domain negative samples to be alternately pushed (intra-domain negative sample redistribution) loss. Triple state loss Where: for each sample b i Select anchor samples from a small sample. negative samples m is the system margin parameter. In this device, m is set to 0.3. + for(·) + =max(·,0), Let K be the average eigenvector of the region K in the mini-batch. S K represents the number of meta-training domains in this system. s Let N be 2. S K N is the number of meta-training samples in domain K. S This represents the total number of all meta-training samples. This formula allows for the alternating pushing of negative samples between and within domains, thereby redistributing them across domains. For Euclidean distance, b i H ,b i n- ,b i n+ These represent anchor samples, inter-domain negative samples, intra-domain negative samples, and g, respectively. S It is the softplus function.

[0041] The update process is as follows: Where: LR represents the change in learning rate.

[0042] The preferred step size range for the learning rate change LR is [0.001, 0.01].

[0043] Based on specific practical experiments, the experiment was conducted using the PyTorch framework and an Intel CPU E5-1650V4 processor.

[0044] The system used a 3.60GHz NVIDIA GeForce RTX 2080 Ti GPU and an Ubuntu 16.04 software environment. ResNet50 was selected as the backbone network, using pre-trained weights from ImageNet. The input image was 256×128 pixels. During training, random horizontal flipping, random cropping, and normalization enhancement strategies were employed. The feature learning network was optimized using Stochastic Gradient Descent (SGD). Under specific optimization settings, the momentum was set to 0.9, and the initial learning rate was set to 0.01. The training epochs for both source and target domain pedestrian images were set to 120, with 48 samples per batch. For the first 50 epochs, the ResNet50 network layers were trained using a learning rate of 0.01. In the subsequent 40 epochs, the learning rate was scaled by 10e-1 at each stage, resulting in a learning rate of 0.001.

[0045] Compared with other methods, firstly, the style variation in a single dataset is relatively small, and some pedestrian identities are distributed in different domains. Since the number of domains is different from the previous one, the entire camera domain of each dataset is divided into two parts for training through a meta-learning framework. Secondly, some methods use local attention mechanisms to learn and distinguish partial feature representations, while this invention uses two fully connected (FC) layers with non-linearity to learn global features using global max pooling features. The features are further embedded through 1×1 shallow convolutions, and the triplet loss is input into the meta-training and meta-testing losses to enhance global compactness.

[0046] Tables 1-5 show that, using Rank-k (accuracy) and mAP (mean average accuracy) as evaluation metrics, the proposed method significantly improves upon other methods, demonstrating that the meta-graph-aware approach is superior to generalized pedestrian re-identification methods in related fields. The proposed scheme is clearly better than existing algorithms, exhibiting substantial improvements in both robustness and accuracy, and has significant reference value for real-world scenarios such as target tracking and intelligent surveillance. This method achieved good results on three datasets used for pedestrian re-identification, and the meta-graph-aware design plays a significant role in unsupervised pedestrian re-identification, as shown in Table 6.

[0047] Table 1 shows the comparison experiment of person re-identification across Market-1501 and Market-1501 datasets.

[0048]

[0049] Table 2 shows the comparison experiment of pedestrian re-identification across the DukeMTMC-ReID and Market-1501 datasets.

[0050]

[0051]

[0052] Table 3 shows the comparison experiment of pedestrian re-identification across DukeMTMC-ReID and DukeMTMC-ReID datasets.

[0053]

[0054] Table 4 shows the comparative experiments of person re-identification across the Market-1501 → DukeMTMC-ReID datasets.

[0055]

[0056] Table 5 shows the comparative experiments of person re-identification across datasets: Market-1501→MSMT17 and DukeMTMC-ReID→MSMT17.

[0057]

[0058] Table 6 shows the comparative experimental results of different methods.

[0059]

[0060] As shown in Tables 2, 4, and 5, compared with existing technologies, this method achieves more generalized results in cross-domain person re-identification based on a two-stage approach of meta-global relevance perception and meta-graph relation sampling, solving the problem of insufficient discrimination accuracy caused by inter-domain differences. The research results show that when meta-learning and graph relation sampling-global relevance perception are combined, the performance is improved. Separating the training set helps to significantly improve accuracy. The meta-training loss of this invention is complementary and has a synergistic effect with the recurrent update method. Domain sampling between meta-training and meta-testing is fully balanced. In the meta-learning channel, updating the balance parameters brings stability to the overall training and improves generalization ability. The design of meta-graph perception plays a significant role in domain-generalized person re-identification, with accuracy far exceeding other domain-generalized person re-identification methods. The meta-global perception module and the meta-graph relation sampling module can suppress noise and narrow inter-domain relationships, balancing learning efficiency while preventing model overfitting.

[0061] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A domain-generalized person re-identification method based on meta-graph perception, characterized in that, By constructing a meta-learning framework and inputting source domain training set images into the meta-global relevance perception module of the meta-learning framework without using any pedestrian identity annotations during the meta-training domain stage, a global knowledge graph is constructed based on the obtained pedestrian features, and the distance between these pedestrian features and cluster centers is calculated to construct a pedestrian feature relationship graph between adjacent classes. The pedestrian feature relationship graph is input into the meta-graph relationship sampling module, and features are learned to discriminate between adjacent classes. Then, the optimizer is used to optimize and calculate the meta-training loss function to obtain the pre-trained model. In the meta-test domain stage, the target domain dataset is input into the spatial perception unit and the channel perception unit to obtain the target domain data features and calculate the meta-test loss function. In the online phase, pedestrian images generated by the trained meta-learning framework are used to train a deep neural network for pedestrian re-identification in an unsupervised manner. The model trained in the source domain is used to verify the generalization effect of the model in the target domain. That is, the generalization ability is reflected by the accuracy in the target domain. Four domain generalization pedestrian re-identification benchmark experiments are set up for comparison. The Market1501 and DukeMTMC-ReID datasets are used alternately as the source and target domains for training and testing.

2. The domain-generalized person re-identification method based on meta-graph perception according to claim 1, characterized in that, The global knowledge graph is constructed as follows: in the meta-training domain, all features extracted by the meta-global relevance perception module are processed in blocks, stacked to construct paired relations, and then the paired relations are learned through shallow convolutional model units to construct the global knowledge graph.

3. The domain-generalized person re-identification method based on meta-graph perception according to claim 1, characterized in that, The pedestrian feature relationship graph is constructed as follows: at the beginning of each epoch, a graph is constructed for all pedestrians, and the nearest neighboring class is always sampled in a mini-batch for discriminative learning. The distance from all pedestrian features to the cluster center of the various data feature graphs extracted from the deep neural network of the input global correlation perception mechanism is calculated. The similarity of pedestrian features between domains is used to construct the pedestrian feature relationship graph between adjacent pedestrians.

4. The domain-generalized person re-identification method based on meta-graph perception according to any one of claims 1-3, characterized in that, specifically include: Step 1: Randomly divide the source domain training set images into a meta-training domain and a meta-test domain; Step 2, Meta-training Domain Stage: Using a pre-trained ResNet50 network as the backbone network, the source domain training set images are input into the meta-global correlation perception module to extract various pedestrian features. Based on the obtained pedestrian features, a global knowledge graph is constructed, and the distance from the pedestrian features to the cluster center is calculated to construct a pedestrian feature relationship graph between adjacent classes. This graph is then input into the meta-graph relationship sampling module to sample and learn discriminative features for adjacent classes. Finally, the model is optimized by an optimizer to obtain the pre-trained model. Step 3: Use convolution to reduce dimensionality. Specifically, use a global max pooling layer to obtain feature vectors. Each channel forms a C-dimensional feature map. Reduce the dimensionality of the 2048-dimensional feature map to 512-dimensional feature map by using a 1*1 convolution kernel. Step 4: Input the feature map obtained in Step 3 into the meta-global correlation sensing module containing channel sensing units and spatial sensing units to obtain channel sensing feature node values ​​and spatial sensing feature node values, specifically including: 4.1) The feature space information after average pooling and max pooling operations is input into the shared network to generate the channel sensing node matrix; 4.2) Perform matrix multiplication between the channel sensing node matrix and the original features to obtain the multiplied result matrix; 4.3) Perform element-wise summation on the multiplication result matrix and the original features to obtain the channel-aware feature node values, specifically: ,in: and Implemented by 1×1 convolution and normalization. Convert the channel size to 1; 4.4) After the spatial sensing unit performs two pooling operations, average pooling and max pooling, it concatenates the obtained features and then performs... Convolution, the spatially aware unit is used to learn a size of A spatial diagram, in which: The height of the spatial diagram, The width of the spatial map; 4.5) The average pooling features and max pooling features are concatenated using a 1*1 convolutional layer to obtain the spatially aware feature node values, specifically: ,in: For global flat pooling operations along the channel dimension, Embedding function for the feature itself, For global relation embedding functions, and They are all through space Convolution implementation, for function; Step 5: Superimpose the channel-aware feature node values ​​and spatial-aware feature node values ​​in pairs to obtain the relation-aware node. Used to represent global structure information, where: The operation is a 1×1 convolution; then a global knowledge graph is constructed by learning through shallow convolutional model units. Step 6: Based on the global knowledge graph and the pedestrian features obtained in Step 2, calculate the distance from all pedestrian features to the cluster center, and use this distance to construct a pedestrian feature relationship graph between adjacent classes; Step 7: Perform positive sample cyclic relationship sampling and negative sample random relationship sampling on the pedestrian feature relationship graph between adjacent classes; Step 8: Calculate the loss function: Use in-domain scattering loss and triplet loss in the meta-training loss, and use triplet loss in the meta-testing loss to balance the overall parameter changes, finally obtaining the total meta-training loss and the total meta-testing loss. Update the overall parameters based on the total loss, starting from the meta-training domain. Sampling a small sample ,from Updated to In the meta-test domain Unknown class samples Above, the triplet loss is used to balance the changes in the overall parameters, and the overall parameters are updated meta-updated, specifically as follows: .

5. A meta-graph perception-based domain-generalized pedestrian re-identification system that implements the method of any one of claims 1-4. The system comprises: a ResNet50 network as the backbone network, a meta-global correlation perception module, a meta-graph relationship sampling module, and a pedestrian matching module, wherein: in the meta-training domain, a pre-trained ResNet50 network is used as the backbone network; the meta-global correlation perception module, in order to reduce inter-domain differences, first passes through a channel perception unit and then a spatial perception unit, constructs an interaction model, i.e., pairwise relationships, by stacking block features through an affinity model, calculates relationship-aware nodes, constructs a global knowledge graph, and enhances feature representation capabilities by weighting structural information and suppressing noise; the meta-graph relationship sampling module samples adjacent classes to learn discriminative features; and the pedestrian matching module calculates matching accuracy and precision based on the loss function results using cumulative matching characteristics and average precision.

6. The domain-generalized person re-identification system based on meta-graph perception according to claim 5, characterized in that, The meta-global relevance perception module includes: a shallow convolutional model unit, a channel perception unit, and a spatial perception unit. The shallow convolutional model unit is a 1*1 convolutional layer used to reduce the dimensionality of the 2048-dimensional feature map to 512 dimensions using a 1*1 convolutional kernel. The channel perception unit selects important channels based on the channel perception feature node values ​​generated in the channel domain, obtaining the results for important channels. The spatial perception unit selects important spatial regions and predicts the most relevant spatial locations based on the spatial perception feature node values ​​generated in the spatial domain, obtaining the results for regions of interest. Meta-graph relationship sampling is then performed based on the obtained important channels and important information regions.

7. The domain-generalized person re-identification system based on meta-graph perception according to claim 5, characterized in that, The affinity model described above is specifically: from node To the node Affinity Model Defined as the dot product affinity in the embedding space, i.e., ,in: , The spatial locations of the feature maps are labeled and assigned as follows: ,Will Each feature node is represented as , Embed functions within the node itself. and It is achieved through 1 × 1 spatial convolutional layers and batch normalization ( )and The two embedded functions implemented are activated, namely ,in: , It is a predefined positive integer that controls the dimensionality reduction rate, where C is the feature dimension. Similarly, it also yields the value from the node. To the node The affinity is , where: using the pair To describe and A two-way relationship between them.

8. The domain-generalized person re-identification system based on meta-graph perception according to claim 5, characterized in that, The interaction model is specifically as follows: for the first The interaction model of a set of feature nodes stacked in a fixed order with all other nodes is as follows: ,in: The affinity matrix represents the interaction model between nodes in the network, i.e., pairwise relationships.

Images