Unsupervised person re-identification method based on self-label refinement deep learning model
By using the self-label refinement deep learning model and symmetric contrast loss function, and using local feature clustering to correct pseudo labels, the problem of noisy labels in unsupervised pedestrian re-identification is solved, and the recognition accuracy of the model in large unlabeled datasets and new data domains is improved.
Patent Information
- Application Number
- CN202210246604.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-14
- Publication Date
- 2025-09-09
- Estimated Expiration
- 2042-03-14
AI Technical Summary
Existing unsupervised person re-identification methods are susceptible to the influence of noisy labels during model training due to the lack of labeled data, resulting in performance degradation. In addition, clustering algorithms find it difficult to ensure that image samples of the same pedestrian are assigned to the same pseudo-label, introducing noisy labels.
By constructing a self-label refinement deep learning model, clustering using local features, correcting pseudo labels, combining memory modules and symmetric contrast loss functions, and optimizing the loss function, the network's robustness to noisy labels is improved.
It effectively alleviates cross-view differences, improves the model's robustness to noisy labels, and enhances the accuracy of person re-identification, especially in large unlabeled datasets and new data domains.
Smart Images
Figure CN114648779B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of pedestrian re-identification, and in particular to an unsupervised pedestrian re-identification method based on a self-label refinement deep learning model. Background Art
[0002] Person re-identification (Pedestrian Re-ID) is a popular research topic in the field of computer vision in recent years. It is an image retrieval problem that uses computer vision technology to determine whether a specific person exists in an image or video. That is, given a monitored pedestrian image, it retrieves images of the pedestrian across devices.
[0003] With the development of science and technology, person re-identification technology has been widely used in intelligent security, video surveillance and other fields. At present, person re-identification has made great breakthroughs in the field of labeled supervision and has demonstrated superior performance. However, since labeling datasets requires high labor costs, and unlabeled data is often very cheap and large in scale, how to expand to large unlabeled datasets and new data domains has attracted more and more attention, and unsupervised person re-identification has emerged. Unsupervised person re-identification is to use unlabeled datasets to learn the feature representation of pedestrians. However, due to the lack of annotation of image data, this person re-identification method based on unsupervised learning often finds it difficult to achieve the retrieval accuracy that can be achieved by supervised learning methods.
[0004] Currently, most popular fully unsupervised person re-identification methods use clustering algorithms to generate pseudo-labels for unlabeled samples, thereby training the model in a supervised manner. However, because images of the same person captured from different camera angles vary significantly, clustering algorithms cannot guarantee that images of the same person will be assigned the same pseudo-label. This inevitably introduces noisy labels, making model training susceptible to label noise and leading to reduced performance. Summary of the Invention
[0005] The problem solved by the present invention is how to correct pseudo labels through local features to alleviate the differences in images of the same pedestrian caused by cross-viewing angles, and how to optimize the loss function to further improve the robustness of the network to noisy labels.
[0006] To solve the above problems, the present invention provides an unsupervised person re-identification method based on a self-label refinement deep learning model, comprising the following steps:
[0007] S1: Obtain unlabeled pedestrian image dataset Where N represents the number of images in the dataset, x i Represents the i-th pedestrian image in the dataset, resizes each image to the same height and width, and performs preprocessing;
[0008] S2: Build a self-labeling refinement deep learning model, input the preprocessed training data into the network, and extract multi-granular features of the image samples; the multi-granular features include global features, upper body features, and lower body features;
[0009] S3: Cluster the extracted multi-granularity features to obtain global pseudo labels, upper body pseudo labels, and lower body pseudo labels;
[0010] S4: Build a memory module based on the clustering results, calculate the consistency matrix and correct the global pseudo-label;
[0011] S5: By calculating the total loss, gradient backpropagation updates the network parameters, momentum updates the memory module parameters, and saves the optimal parameters of the network.
[0012] In the above method, local information is fully utilized and pseudo labels are corrected through local features. Local parts usually occupy a smaller area in the entire image, so the differences in images of the same pedestrian caused by cross-viewing angles can be alleviated. By clustering local features, samples of the same pedestrian from different cameras can be effectively clustered into the same category. Therefore, the pseudo-label noise caused by clustering global features can be effectively corrected based on local results. In unsupervised person re-identification, contrast loss combined with memory modules is often used. In order to better solve the accuracy loss caused by noisy labels on network training, the symmetric contrast loss symmetricInfoNCEloss is used, which swaps the label and classification confidence. Because the noisy pseudo-label cannot represent the true category distribution, the classification confidence can reflect the true category distribution to a certain extent, thereby improving the robustness of the network to noisy labels.
[0013] Furthermore, the preprocessing includes four methods: horizontal flipping, zero padding, normalization and random erasing.
[0014] Furthermore, the self-label refinement deep learning model consists of a feature extractor, a memory module and a self-label refinement module;
[0015] The feature extractor structure consists of two parts. The first part of the feature extractor adopts the first four stages of the ResNet50 architecture. The second part of the feature extractor is divided into two branches after the first part, namely the global branch and the local branch. The feature map obtained by the global branch is pooled using Generalize-meanPooling, where the parameter p=2 and the size of the feature map after pooling is 2048×1×1; the feature map obtained by the local branch is pooled using Generalize-meanPooling, where the parameter p=2 and the size of the feature map after pooling is 2048×2× 1, and divide the feature map into horizontal strips to obtain two local feature maps of size 2048×1×1; reduce the feature map channel obtained by the local branch, use 256 1×1 convolution kernels to convolve the two 2048×1×1 local feature maps respectively, perform BatchNormalize and ReLU on the results, and the local branch obtains two local feature maps of size 256×1×1 and 256×1×1; perform dimension pruning on the feature maps obtained by the global branch and the local branch to obtain a global feature with a dimension of 2048×1 and two local features with a dimension of 256×1.
[0016] Furthermore, the ResNet50 architecture includes five stages:
[0017] Phase 1: Convolution with 64 kernels, 7×7 kernel size, 3 zero padding parameters, and a stride of 2; BatchNormalize and ReLU activation; Max pooling with a 3×3 kernel size, 1 zero padding parameter, and a stride of 2.
[0018] The second stage includes three Bottlenecks;
[0019] The third stage includes four Bottlenecks;
[0020] The fourth stage consists of six Bottlenecks;
[0021] The fifth stage consists of three Bottlenecks.
[0022] Furthermore, the structure of the global branch is the same as the fifth stage of the ResNet50 architecture, including three Bottlenecks, except that the step size of the second convolution operation of the first Bottleneck is 1; wherein, the convolution step size in the downsampling operation is 1.
[0023] Furthermore, the structure of the local branch is the same as the fifth stage of the ResNet50 architecture, including three bottles, except that all convolution steps in the first bottleneck are 1; wherein, the convolution step in the downsampling operation is 1, and all convolution steps in the other two bottles are 1, and the downsampling operation is canceled.
[0024] Furthermore, the pre-processed training data is input into the network so that all pedestrian images in the dataset are input into the feature extractor to obtain the feature set
[0025] in Represents picture x respectively i Global features, upper body features and lower body features.
[0026] Furthermore, the step S4 includes:
[0027] S41: Calculate the Jaccard distance between the global features of all images and obtain an N×N dimensional distance matrix;
[0028] S42: Combine the obtained distance matrix and use DBSCAN to perform clustering. The samples in a cluster are assigned the same pseudo label, and the outliers generated by the cluster are assigned the pseudo label of the nearest cluster.
[0029] S43: For local features, including upper body features and lower body features, the Jaccard distance matrix is calculated respectively, and DBSCAN is used to perform clustering operations on each. For outliers, the method of assigning the nearest cluster pseudo label is also adopted to obtain their respective pseudo labels.
[0030] Furthermore, the calculation of the consistency matrix is as follows: the sample set with the global pseudo label n is recorded as I g (n), where n∈[1,Z g ], the sample set with the upper body pseudo label j and the lower body pseudo label k is recorded as I up (j) and I low (k), where j∈[1,Z up ],k∈[1,Z low ],Z up and Z low are the number of categories obtained by clustering the upper body and lower body respectively;
[0031] The modified global pseudo label is given a sample x i , get its one-hot global pseudo label Upper body features and lower body features Combine the obtained consistency matrix and memory module to calculate the confidence of the transferred local label and correct the global pseudo label;
[0032] The training sample data is randomly divided into multiple batches. Each batch contains P categories, each category has K pictures, and a batch has P×K picture samples. A sample x i It includes three features, namely global features Upper body features and lower body features It also includes three labels, namely global pseudo labels Upper body pseudo-label and lower body pseudo-labels Correct the global pseudo label of the sample and get the corrected label as
[0033] Furthermore, the memory module is updated in the following manner: That is, the agent of class j is updated by the features of the difficult positive samples belonging to class j in the batch, μ is the momentum update coefficient, where μ = 0.2; repeat steps 4 and 5 to train the dataset once, that is, run one epoch; after one epoch, continue to run steps 2 and 3; run num_epoch times, where num_epoch = 50.
[0034] The present invention adopts the above technical solution to achieve at least the following beneficial effects:
[0035] By clustering local features, this method can effectively group samples of the same person from different cameras into the same class. This localized approach effectively corrects the pseudo-label noise caused by clustering global features. Using a symmetric contrast loss, it swaps label and classification confidence. Because noisy pseudo-labels cannot represent the true class distribution, while classification confidence can reflect it to a certain extent, this improves the network's robustness to noisy labels. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 The unsupervised person re-identification method based on the self-label refinement deep learning model provided by the embodiment of the present invention Figure 1 ;
[0037] Figure 2 The unsupervised person re-identification method based on the self-label refinement deep learning model provided by the embodiment of the present invention Figure 2 ;
[0038] Figure 3Schematic diagram of the algorithm operation of the self-label refinement model of the unsupervised pedestrian re-identification method based on the self-label refinement deep learning model provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0039] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.
[0040] The following are specific embodiments of the present invention and the accompanying drawings to further describe the technical solutions of the present invention, but the present invention is not limited to these embodiments.
[0041] Example
[0042] This embodiment provides an unsupervised person re-identification method based on a self-labeled refined deep learning model. Figures 1 to 3 As shown, the method includes the steps of:
[0043] S1: Obtain unlabeled pedestrian image dataset Where N represents the number of images in the dataset, x i Represents the i-th pedestrian image in the dataset, resizes each image to the same height and width, and performs preprocessing;
[0044] S2: Build a self-labeling refinement deep learning model, input the preprocessed training data into the network, and extract multi-granular features of the image samples; the multi-granular features include global features, upper body features, and lower body features;
[0045] S3: Cluster the extracted multi-granularity features to obtain global pseudo labels, upper body pseudo labels, and lower body pseudo labels;
[0046] S4: Build a memory module based on the clustering results, calculate the consistency matrix and correct the global pseudo-label;
[0047] S5: By calculating the total loss, gradient backpropagation updates the network parameters, momentum updates the memory module parameters, and saves the optimal parameters of the network.
[0048] Among them, preprocessing includes four methods: horizontal flipping, zero padding, normalization and random erasing.
[0049] Among them, the self-label refinement deep learning model consists of a feature extractor, a memory module and a self-label refinement module;
[0050] The feature extractor structure consists of two parts. The first part of the feature extractor adopts the first four stages of the ResNet50 architecture. The second part of the feature extractor is divided into two branches after the first part, namely the global branch and the local branch. The feature map obtained by the global branch is pooled using Generalize-meanPooling, where the parameter p=2 and the size of the feature map after pooling is 2048×1×1; the feature map obtained by the local branch is pooled using Generalize-meanPooling, where the parameter p=2 and the size of the feature map after pooling is 2048×2×1. , and divide the feature map into horizontal strips to obtain two local feature maps of size 2048×1×1; reduce the feature map channel obtained by the local branch, use 256 1×1 convolution kernels to convolve the two 2048×1×1 local feature maps respectively, perform BatchNormalize and ReLU on the results, and the local branch obtains two local feature maps of size 256×1×1 and 256×1×1; the feature maps obtained by the global branch and the local branch are dimensionally cropped to obtain a global feature with a dimension of 2048×1 and two local features with a dimension of 256×1.
[0051] Among them, the ResNet50 architecture consists of five stages:
[0052] Phase 1: Convolution with 64 kernels, 7×7 kernel size, 3 zero padding parameters, and a stride of 2; BatchNormalize and ReLU activation; Max pooling with a 3×3 kernel size, 1 zero padding parameter, and a stride of 2.
[0053] The second stage includes three Bottlenecks;
[0054] The third stage includes four Bottlenecks;
[0055] The fourth stage consists of six Bottleneck
[0056] The fifth stage consists of three Bottlenecks.
[0057] The structure of the global branch is the same as the fifth stage of the ResNet50 architecture, including three bottles. The difference is that the stride of the second convolution operation of the first bottleneck is 1; the convolution step in the downsampling operation is 1.
[0058] The structure of the local branch is the same as the fifth stage of the ResNet50 architecture, including three bottles. The difference is that all convolution steps in the first bottleneck are 1; the convolution step in the downsampling operation is 1, and all convolution steps in the other two bottles are 1, and the downsampling operation is canceled.
[0059] The pre-processed training data is input into the network. All pedestrian images in the dataset are input into the feature extractor to obtain the feature set.
[0060] in Represents picture x respectively i Global features, upper body features and lower body features.
[0061] See Figure 2 , wherein step S4 includes:
[0062] S41: Calculate the Jaccard distance between the global features of all images and obtain an N×N dimensional distance matrix;
[0063] S42: Combine the obtained distance matrix and use DBSCAN to perform clustering. The samples in a cluster are assigned the same pseudo label, and the outliers generated by the cluster are assigned the pseudo label of the nearest cluster.
[0064] S43: For local features, including upper body features and lower body features, the Jaccard distance matrix is calculated respectively, and DBSCAN is used to perform clustering operations on each. For outliers, the method of assigning the nearest cluster pseudo label is also adopted to obtain their respective pseudo labels.
[0065] It is worth noting that when calculating the local distance matrix, we use the features of local features and global features to calculate the similarity. After the above process, we can get a new dataset That is, picture x i It has three pseudo labels, which come from the clustering results of global features. Clustering results of upper body features Clustering results of lower body features For the clustering results of global features, the average value of sample features in each category is calculated as the feature center of each category, and memory space is opened for storage to obtain where Z g Indicates the number of categories obtained by global feature clustering, and 2048 indicates the dimension of the average feature vector. In the same way, we can get and where Z up and Z lowThey represent the number of categories obtained by clustering the upper and lower body features, and 256 represents the dimension of each type of central feature of the upper and lower body.
[0066] See Figure 3 , where the consistency matrix is calculated as follows: the sample set with global pseudo label n is recorded as I g (n), where n∈[1,Z g ], the sample set with the upper body pseudo label j and the lower body pseudo label k is recorded as I up (j) and I low (k), where j∈[1,Z up ],k∈[1,Z low ],Z up and Z low are the number of categories obtained by clustering the upper body and lower body respectively;
[0067] Correct the global pseudo label to be given a sample x i , get its one-hot global pseudo label Upper body features and lower body features Combine the obtained consistency matrix and memory module to calculate the confidence of the transferred local label and correct the global pseudo label;
[0068] The training sample data is randomly divided into multiple batches. Each batch contains P categories, each category has K pictures, and a batch has P×K picture samples. A sample x i It includes three features, namely global features Upper body features and lower body features It also includes three labels, namely global pseudo labels Upper body pseudo-label and lower body pseudo-labels Correct the global pseudo label of the sample and get the corrected label as
[0069] Specifically, for the calculation method of cluster consistency matrix, For example, the consistency between the upper body class j and the global class i in the matrix is calculated by IOU, and the formula is where |·| represents the number of samples in the set. We use the same method to calculate Normalize the two consistency matrices, that is, the sum of each row value in the matrix is 1, so we get the standardized matrix and
[0070] Specifically, calculate the local label confidence, multiply the local feature and the corresponding memory module to obtain the label confidence and The dimensions are Z up ×1 and Z low ×1. Since the categories of local labels and global labels are not unified, it is impossible to directly transfer local information to global labels. Using the consistency matrix, we can cleverly map the local label confidence to the corresponding global pseudo label, that is, to obtain the transferred local confidence and in τ is the temperature coefficient hyperparameter, with a value of 10, which plays a scaling role, Τ represents the transpose of the matrix, and softmax is the normalized exponential function. The final sample x i The global label is corrected to Where α∈[0,1] is the momentum coefficient for updating labels and decays exponentially α=α0*(1-now_epoch / num_epoch) 0.6 , where now_epoch is the current epoch of training, num_epoch represents the total number of training epochs, and α0 is the initial momentum coefficient, which is 0.2.
[0071] Specifically, there are three ways to calculate the total loss:
[0072] Contrast loss InfoNCE loss, the formula is Where f is the query feature, and the clustering result shows that it belongs to the jth class. is the memory module corresponding to the query feature, is the positive sample proxy of the query feature in the memory module, τ is the temperature coefficient, <·,·> represents the inner product operation between two features, which measures the similarity between the two features, Z is the number of corresponding categories obtained by clustering, and exp(·) represents the exponential function with the natural constant e as the base.
[0073] Symmetric InfoNCE loss, the formula is Where y is the one-hot pseudo-label corresponding to f, and softmax(y)[t] represents the value of the t-th dimension after the pseudo-label is normalized by the exponential function.
[0074] Soft softmax-triplet loss, the formula is specific in represents the binary cross entropy loss, f + and f -Respectively represent the difficult positive sample features and difficult negative sample features relative to f in the batch. + Indicates the label corresponding to the difficult positive sample, y - Denotes the labels corresponding to the hard negative samples. ||·|| denotes L2 normalization.
[0075] Based on these three losses, the sample x i The loss is divided into global loss and local losses Total loss Its λ is the balance coefficient, in this invention λ = 0.5, it is worth noting that in the global loss middle, and The corrected soft label used is The final batch loss is Where B represents the batch set and i is a sample in the batch.
[0076] Among them, the memory module update method is That is, the agent of class j is updated by the features of the difficult positive samples belonging to class j in the batch, μ is the momentum update coefficient, where μ = 0.2; repeat steps 4 and 5 to train the dataset once, that is, run one epoch; after one epoch, continue to run steps 2 and 3; run num_epoch times, where num_epoch = 50.
[0077] The deep learning network used in this method was implemented using the Pytorch framework and ran on a Quadro P5000. The Adam optimizer was used, with a learning rate of 0.00035 and a training iteration count of Num_epoch = 50. Before training, images were resized to 256×128 pixels, and the batch size was 64, containing 16 pseudo-classes with 4 samples per class. Recognition performance was measured using two metrics: mAP and Rank-1.
[0078] Ablation experiments: To evaluate the superiority of the present invention, the proposed self-label refinement module and symmetric information loss function are evaluated. The results are verified on three public person re-identification datasets, as shown in Table 1:
[0079]
[0080] Table 1
[0081] The first row in Table 1 shows the base network, which is the existing basic model for unsupervised person re-identification. The second row shows the results after adding the Self-Label Refining Module (SLR) to the base network. It can be seen that the addition of the module improves the network model performance. The third row of experimental results shows that adding symmetric contrast loss to the base network also achieves a stable improvement in model performance. The last row shows the experimental results of the model after adding both the Self-Label Refining Module and the Symmetric Information Loss to the base network model architecture. Compared with the previous experiments, this experimental result reaches the best.
[0082] Comparative Experiments: To evaluate the effectiveness of the self-label refinement network in our method, we compared it with several other existing unsupervised person re-ID models, including the unsupervised domain adaptation (UDA) person re-ID method and the fully unsupervised (FU) person re-ID method. The results of the comparative experiments are shown in Table 2:
[0083]
[0084]
[0085] Table 2
[0086] Because unsupervised domain adaptation (UDA) methods introduce a labeled source domain, they generally outperform fully unsupervised (FU) methods. This method, without using labeled data, approaches the performance of domain adaptation methods. Table 2 shows that this method achieves good results on three public person re-identification datasets: Market-1501, DukeMTMC-reID, and MSMT17. This demonstrates the superiority of our proposed self-label refinement strategy and symmetric information loss.
[0087] By clustering local features, this method can effectively group samples of the same person from different cameras into the same class. This localized approach effectively corrects for pseudo-label noise caused by clustering global features. Using a symmetric contrast loss, it swaps label and classification confidence, as noisy pseudo-labels cannot represent the true class distribution, while classification confidence can reflect it to a certain extent. This improves the network's robustness to noisy labels.
[0088] Although the present disclosure is disclosed as above, the protection scope of the present disclosure is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present disclosure, and these changes and modifications will fall within the protection scope of the present invention.
Claims
1. Unsupervised person re-identification method based on self-label refinement deep learning model, characterized by: Including steps: S1: Obtain unlabeled pedestrian image dataset Where N represents the number of images in the dataset, x i Represents the i-th pedestrian image in the dataset, resizes each image to the same height and width, and performs preprocessing; S2: Build a self-labeling refinement deep learning model, input the preprocessed training data into the network, and extract multi-granular features of the image; the multi-granular features include global features, upper body features, and lower body features; The self-label refinement deep learning model consists of a feature extractor, a memory module and a self-label refinement module; The feature extractor structure consists of two parts. The first part of the feature extractor adopts the first four stages of the ResNet50 architecture. The second part of the feature extractor is divided into two branches after the first part, namely the global branch and the local branch. The feature map obtained by the global branch is pooled using Generalize-meanPooling, where the parameter p=2 and the size of the feature map after pooling is 2048×1×1; the feature map obtained by the local branch is pooled using Generalize-meanPooling, where the parameter p=2 and the size of the feature map after pooling is 2048×2× 1, and divide the feature map into horizontal strips to obtain two local feature maps of size 2048×1×1; reduce the feature map channel obtained by the local branch, use 256 1×1 convolution kernels to convolve the two 2048×1×1 local feature maps respectively, perform BatchNormalize and ReLU on the obtained results, and the local branch obtains two local feature maps of size 256×1×1 and 256×1×1; perform dimension clipping on the feature maps obtained by the global branch and the local branch to obtain a global feature with a dimension of 2048×1 and two local features with a dimension of 256×1; S3: Cluster the extracted multi-granularity features to obtain global pseudo labels, upper body pseudo labels, and lower body pseudo labels; S4: Build a memory module based on the clustering results, calculate the consistency matrix and correct the global pseudo-labels; the specific process is as follows: S41: Calculate the Jaccard distance between the global features of all images and obtain an N×N dimensional distance matrix; S42: Based on the obtained distance matrix, DBSCAN is used for clustering. The images in a cluster are assigned the same pseudo label, and the outliers generated by the cluster are assigned the pseudo label of the nearest cluster. S43: For local features, including upper body features and lower body features, the Jaccard distance matrix is calculated respectively, and DBSCAN is used to perform clustering operations on each. For outliers, the method of assigning the nearest cluster pseudo label is also adopted to obtain their respective pseudo labels; S5: By calculating the total loss, backpropagating the gradient to update the network parameters, and using momentum to update the memory module parameters, the optimal parameters of the network are saved; The ResNet50 architecture consists of five stages: Phase 1: Convolution with 64 kernels, 7×7 kernel size, 3 zero padding parameters, and a stride of 2; Batch Normalization and ReLU activation; Max Pooling with a 3×3 kernel size, 1 zero padding parameter, and a stride of 2. The second stage includes three Bottlenecks; The third stage includes four Bottlenecks; The fourth stage consists of six Bottlenecks; The fifth stage consists of three Bottlenecks; The pre-processed training data is input into the network. All pedestrian images in the dataset are input into the feature extractor to obtain the feature set. in Represents picture x respectively i Global features, upper body features and lower body features; The calculation consistency matrix is as follows: the set of pictures with global pseudo labels n is recorded as I g (n), where n∈[1,Z g ],Z g It represents the number of categories obtained by global feature clustering. The set of pictures with the upper body pseudo label j and the lower body pseudo label k is recorded as I up (j) and I low (k), where j∈[1,Z up ],k∈[1,Z low ],Z up and Z low are the number of categories obtained by clustering upper body features and lower body features; The modified global pseudo label is given a picture x i , get its one-hot global pseudo label Upper body features and lower body features Combine the obtained consistency matrix and memory module to calculate the confidence of the transferred local label and correct the global pseudo label; The training image data is randomly divided into multiple batches. Each batch contains P categories, each category has K pictures, and a batch has P×K pictures. i It includes three features, namely global features Upper body features and lower body features It also includes three labels, namely global pseudo labels Upper body pseudo-label and lower body pseudo-labels Correct the global pseudo-label of the image and get the corrected label:
2. The unsupervised person re-identification method based on self-label refinement deep learning model according to claim 1 is characterized in that: The preprocessing includes four methods: horizontal flipping, zero padding, normalization and random erasing.
3. The unsupervised person re-identification method based on self-label refinement deep learning model according to claim 1 is characterized in that: The structure of the global branch is the same as the fifth stage of the ResNet50 architecture, including three Bottlenecks, except that the step size of the second convolution operation of the first Bottleneck is 1; wherein, the convolution step size in the downsampling operation is 1.
4. The unsupervised person re-identification method based on self-label refinement deep learning model according to claim 1 is characterized in that: The structure of the local branch is the same as the fifth stage of the ResNet50 architecture, including three bottles. The difference is that all convolution steps in the first bottleneck are 1; among them, the convolution step in the downsampling operation is 1, and all convolution steps in the other two bottles are 1, and the downsampling operation is canceled.