CN114898406A - Unsupervised pedestrian re-identification method based on contrast clustering - Google Patents

Abstract

The invention discloses an unsupervised pedestrian re-identification method based on contrast clustering, which comprises the following steps: performing forward calculation on the unmarked pedestrian image data set by using an initial feature encoder, and initializing a feature storage unit by using the encoded features; clustering the feature storage units before each round of training, and screening clustering results according to a clustering independence standard and a clustering closeness standard; carrying out feature coding on each group of small-batch training samples, and carrying out backward propagation updating on the network by using a uniform contrast loss function; dynamically updating instance features in the feature storage unit in a momentum updating manner by using the encoded features; and circularly updating the feature encoder and the feature storage unit according to the preset number of training rounds until the pedestrian re-identification network converges. The invention fully excavates the available information of the non-clustered outlier by using a contrast clustering method, and improves the identification accuracy of the unsupervised pedestrian re-identification model.

Description

An Unsupervised Person Re-ID Method Based on Contrastive Clustering

technical field

The invention relates to the fields of computer vision and pedestrian re-identification, in particular to an unsupervised pedestrian re-identification method based on contrastive clustering.

Background technique

Pedestrian re-identification, also known as person re-identification, is considered as a sub-problem of image retrieval, with the goal of retrieving specific pedestrians in multiple non-overlapping surveillance camera regions. Pedestrian re-identification technology can make up for the visual limitations of a single fixed camera, and can be combined with pedestrian detection and pedestrian tracking technology, and is widely used in intelligent monitoring, video tracking and other security fields. With the development of deep learning technology and the introduction of large-scale datasets, the current supervised person re-identification methods have achieved a huge improvement in performance. However, algorithms based on supervised learning greatly rely on manually annotated real labels, hindering The further development of pedestrian re-identification technology. On the other hand, it is easy to obtain a large amount of unlabeled pedestrian image data in reality, and it is of great research value to study how to use large-scale unlabeled pedestrian images to train a more robust person re-identification model. Therefore, an unsupervised person re-identification method without any annotation information is proposed to solve the above problems.

Unsupervised person re-identification methods mainly include methods based on pseudo-labels and methods based on image generation. Among them, the pseudo-label method based on clustering has been proved to be more effective and maintains the current state-of-the-art accuracy. At present, most of the clustering-based pseudo-label methods are divided into two steps in training: the first step is to use the initial feature encoder to encode pedestrian images; the second step is to cluster the encoded features to obtain pseudo labels, as Supervise the training of the network. Although this kind of method can continuously improve the quality of pseudo-labels with the optimization of the model to a certain extent, the training of the model is often disturbed by the unavoidable pseudo-label noise, and when the initial pseudo-label noise is large, the model has more Great risk of collapse. In addition, cluster-based pseudo-label methods often do not use all unlabeled training data, and density-based clustering algorithms themselves will generate cluster outliers. These cluster outliers cannot be assigned pseudo-labels. are discarded and not used for model training. However, such clustering outliers are often the difficult training samples worth mining in the pedestrian dataset. Especially in the early stage of training, there are often a large number of clustering outliers. Simply discarding them will greatly reduce the training samples and seriously damage the model. performance.

Contrastive learning has been widely used in unsupervised representation learning tasks in recent years. Contrastive learning can fully learn the similarity between samples of the same category and the differences between different categories without supervision. Each unlabeled sample is regarded as a different class, and the discriminative representation of the sample instance is learned by optimizing the contrast loss. However, most of the current contrastive losses are at the sample instance level, and it is difficult to correctly measure the intra-class relationship in pedestrian image datasets.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present invention proposes an unsupervised pedestrian re-identification method based on contrastive clustering. Through joint contrastive learning of cluster centroids and unclustered outliers, the difficult training in the target domain data set can be fully exploited. samples, and effectively model the intra-class relationships of pedestrians.

Specifically, the technical solution of an unsupervised pedestrian re-identification method based on contrastive clustering proposed by the present invention includes the following steps:

Step 1: Use the initial feature encoder f _θ to perform forward calculation on the pedestrian images in the unlabeled training data set, and use the encoded features to initialize the feature storage unit based on the category prototype;

Step 2: Before each round of training, use the DBSCAN clustering algorithm to cluster the coding features in the feature storage unit, and screen the clustering results according to the clustering reliability evaluation standard;

Step 3: For each group of mini-batch training samples, use the encoder f _θ to perform feature encoding on them to obtain the mini-batch sample feature f, and use the unified contrast loss function to calculate the loss between the mini-batch sample feature f and the features in the feature storage unit , and perform back-propagation updates to the network;

Step 4: In each iterative training process, the feature storage unit is dynamically updated in the manner of momentum update by using the coding features obtained by forward calculation of the training samples in small batches;

Step 5: According to the preset number of training rounds, repeat steps 2-4 until the pedestrian re-identification model converges.

Further, the initialization process of the feature encoder and feature storage unit in step 1 is:

Use a ResNet-50 deep neural network as the feature encoder f _θ and initialize it with pretrained weights on the ImageNet image dataset;

Use the feature encoder f _θ to perform forward calculation on the samples in the pedestrian image data set to extract features, and obtain a feature set {v ₁ ,...,v _n }, where n represents the number of samples in the pedestrian image data set, and the features in the feature set are All instances are stored in the feature storage unit, so that the category prototypes in the feature storage unit can still be continuously updated when the clustered clusters and unclustered outliers are constantly changing.

Further, the clustering and screening process of the features in step 2 is:

及未聚类离群值实例

其中n_c表示聚类簇的数量，n_o表示未聚类离群值的数量，随后依据聚类独立性与聚类紧密性标准，对聚类的结果进行筛选，并采用k-reciprocal近邻算法对检索结果进行重排序。First, the DBSCAN clustering algorithm is used to cluster the feature set {v ₁ ,...,v _n } in the feature storage unit in step 1, and the category prototypes in the feature storage unit are further divided into cluster centroids

and unclustered outlier instances

where n _c represents the number of clustered clusters, and n _o represents the number of unclustered outliers. Then, according to the criteria of cluster independence and cluster tightness, the clustering results are screened, and the k-reciprocal nearest neighbor algorithm is used. Reorder search results.

Both clustered clusters and unclustered instances in the feature storage unit are regarded as equal and independent classes, so the reliability of clustering is very important to the impact of training, and the network is more discriminative for images at the beginning of training. poor, and the noise of the clustering is also large, so a self-step learning strategy is proposed to improve the effect of clustering. Specifically, re-clustering is performed before the start of each round of training, starting from the most reliable clusters, retaining the reliable clusters, and disassembling the features in the unreliable clusters back to cluster-free outliers In the example, the number of clusters is gradually increased, and more reliable clustering results are obtained by adjusting the ε-neighbor distance threshold of the samples in the DBSCAN clustering algorithm, and alternately relaxing and tightening the clustering criteria.

The cluster independence criterion is used to measure the distance between classes, which is expressed as the intersection ratio between the feature set and the feature set after relaxing the clustering criteria:

where |·| represents the number of features in the set, I( _fi ) represents the sample set in the same cluster, I _loose ( _fi ) represents the sample set in the same cluster after relaxing the clustering criteria, R _indep ( f _i ) represents the independence score of cluster I(fi ₎ ;

The cluster tightness criterion is used to measure the intra-class distance, which is expressed as the intersection ratio between the feature set and the feature set after the clustering criterion is compressed:

where I _tight ( _fi ) represents the sample set in the same cluster after the clustering standard is tightened, and R _comp ( _fi ) represents the tightness score of the cluster I ( _fi );

The above cluster reliability evaluation criteria are used to measure the independence between clusters and the closeness between samples. The starting point of the cluster reliability evaluation criteria is that a reliable cluster should be based on multi-scale clustering. Stable under the environment, set the hyperparameter α,β∈[0,1] to represent the threshold of independence and compactness, keep the inter-class independence R _comp (f _i )>α and the intra-class compactness R _indep (f _i )>β's clustered cluster samples, and the remaining samples are divided into unclustered outliers.

Further, the unified contrastive loss function in step 3 is expressed as:

通过特征编码器编码后将特征全部存储到特征存储单元中，使用步骤2中的自步学习策略将特征集合划分为聚类簇特征和未聚类离群值特征，从而整个训练数据集被分为有聚类伪标签的样本集合

和不属于任何聚类的离群值实例样本集合

且

Given unlabeled training samples

After encoding through the feature encoder, all the features are stored in the feature storage unit, and the feature set is divided into clustered cluster features and unclustered outlier features using the self-paced learning strategy in step 2, so that the entire training data set is classified into is a set of samples with clustered pseudo-labels

and a sample set of outlier instances that do not belong to any cluster

and

使用特征编码器对每个训练样本进行前向计算得到编码的特征f，构建统一对比损失函数：Given training samples

Use the feature encoder to perform forward calculation on each training sample to obtain the encoded feature f, and construct a unified contrastive loss function:

Among them, z ⁺ represents the positive category prototype of the feature f, τ represents the temperature coefficient, <·,·> represents the vector inner product, _ck is the centroid of the current cluster k, which represents the category prototype in the cluster, and v _k is the current cluster. Instance features of class k outliers, representing class prototypes without clustering;

If f belongs to cluster k, then z ⁺ =c _k is the centroid of cluster k; if f belongs to unclustered outliers, then z ⁺ =v _k is the instance feature of unclustered outliers. The above comparison The loss promotes encoding features close to their true classes. After encoding features in mini-batches, they are compared with two class prototypes, so that each training sample is close to the class it belongs to and far away from other classes.

Further, the process of updating the momentum of the feature storage unit in step 4 is:

First, the training samples are all stored in the instance unit, and then the features in each mini-batch sample are accumulated into the instance features corresponding to the feature storage unit in the form of momentum update according to the index number;

而未聚类离群值的实例特征

直接从特征存储单元中提取剩余的实例特征，第k个聚类簇的质心表示为：In the feature storage unit, the average value between the features is calculated for the features {v ₁ ,...,v _n } in the same cluster to obtain the cluster centroid

while the instance features of the unclustered outliers

The remaining instance features are directly extracted from the feature storage unit, and the centroid of the kth cluster is expressed as:

where I _k represents the feature vector set of the kth cluster, |·| represents the number of feature vectors in the set, and the instance feature {v} in the feature storage unit is initially initialized once by the network forward calculation, and in the subsequent training process is continuously updated for more robust clustering;

In each round of training, use the encoded features in the small batch of samples to update the category prototype in the feature storage unit in a momentum update manner, and perform a momentum weighted summation on the instance feature v _i and the current sample feature f _i to obtain the updated Instance features of :

v _i ←mv _i +(1-m)f _i

where m∈[0,1] is the momentum factor, which represents the proportion of instance feature vi and sample feature _fi in the momentum update process. Given the updated vi _, if _fi belongs to cluster _k , then The corresponding cluster centroids _ck are updated accordingly.

Further, the overall training process of this method in step 5 is:

通过ResNet-50特征编码器进行特征编码，使用得到的特征集合初始化特征存储单元；For unlabeled training datasets

The feature encoding is performed by the ResNet-50 feature encoder, and the feature storage unit is initialized with the obtained feature set;

划分为聚类簇样本

和无聚类离群值样本

计算

中的聚类质心；In each subsequent training round, the feature set {v} in the feature storage unit is clustered according to the cluster independence and cluster tightness criteria, and the training samples are

Divide samples into clusters

and cluster-free outlier samples

calculate

The cluster centroids in ;

For training samples in each mini-batch, use the feature encoder f _θ for feature encoding, compute the contrastive loss, and perform backpropagation to update the encoder f _θ ;

Update the feature set {v} in the feature storage unit, and then update the cluster centroid _ck according to the updated feature set {v};

The feature encoder f _θ and feature storage unit are updated cyclically until the model achieves better convergence.

Compared with the prior art, the beneficial effects of the present invention are:

(1) The present invention proposes that the feature storage unit stores all valid information on the unlabeled data domain to learn more sufficient feature representations, and provides two kinds of supervision information of clustering and unclustered outlier instances by dynamically updating the feature storage unit, which is sufficient Utilize difficult training samples from unlabeled datasets;

(2) The present invention proposes to use the self-step learning strategy to screen the clustering results. The network training process starts from the most reliable clustering in the feature storage unit, and puts more unclustered outliers into new clustering clusters. Perform multiple feature clustering in the process to obtain more reliable clustering results, and gradually improve the discriminativeness of feature representation, which can effectively alleviate the problem of pseudo-label noise and optimize the learning process of feature representation;

(3) The present invention proposes a multi-scale clustering reliability measurement standard, including clustering independence and clustering tightness standards, starting from the most reliable clustering cluster, and then gradually increasing the number of clustering clusters, from simple to difficult , combined with the self-step learning strategy to gradually create more reliable clusters, realize the dynamic optimization of the feature storage unit and the person re-identification network, and greatly improve the performance of the unsupervised person re-identification model.

Description of drawings

1 is a flowchart of an unsupervised pedestrian re-identification method based on contrast clustering according to an embodiment of the present invention;

2 is a system schematic diagram of an unsupervised pedestrian re-identification method based on comparative clustering according to an embodiment of the present invention;

3 is a schematic diagram of a unified comparison loss calculation according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of momentum update of a characteristic storage unit according to an embodiment of the present invention.

specific implementation

The specific technical solutions of the present invention will be described in further detail below with reference to the embodiments and the accompanying drawings. The following embodiments or drawings are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

System embodiment

Referring to FIG. 1 to FIG. 4 , the present embodiment provides an unsupervised pedestrian re-identification method based on contrastive clustering, and the method includes the following steps:

Step 1: Use the initial feature encoder f _θ to perform forward calculation on the pedestrian images in the unlabeled training data set, and use the encoded features to initialize the feature storage unit based on the category prototype;

Step 2: Before each round of training, use the DBSCAN clustering algorithm to cluster the coding features in the feature storage unit, and screen the clustering results according to the clustering reliability evaluation standard;

Step 3: For each group of mini-batch training samples, use the encoder f _θ to perform feature encoding on them to obtain the mini-batch sample feature f, and use the unified contrast loss function to calculate the loss between the mini-batch sample feature f and the features in the feature storage unit , and perform back-propagation updates to the network;

Step 4: In each iterative training process, the feature storage unit is dynamically updated in the manner of momentum update by using the coding features obtained by forward calculation of the training samples in small batches;

Step 5: According to the preset number of training rounds, repeat steps 2-4 until the pedestrian re-identification model converges.

The above steps can be summarized as the following processes: (1) initialization process of feature encoder and feature storage unit; (2) feature clustering and screening process; (3) unified contrast loss calculation process; (4) feature storage unit momentum update process; (5) the overall network training process.

A detailed description will be given below.

(1) Initialization process of feature encoder and feature storage unit

Use a ResNet-50 deep neural network as the feature encoder f _θ and initialize its parameters with pretrained weights on the _ImageNet image dataset, replacing the fully connected layer of the last layer with a batch normalization layer and an L2 regularization layer , to accommodate the needs of unsupervised tasks;

During training, each mini-batch sample contains 64 unlabeled pedestrian images belonging to at least 16 different classes. In the case where clustered clusters and unclustered outliers are treated as separate classes, each cluster Clusters are assigned 4 pedestrian images, and each unclustered outlier is assigned a single pedestrian image. The size of each pedestrian image in the dataset is resized to 256×128 before training, and random flipping, random cropping, and random erasing are used. Data augmentation is performed by means of division and so on.

Use the feature encoder f _θ to perform forward calculation on the samples in the pedestrian image data set to extract features, and obtain a feature set {v ₁ ,...,v _n }, where n represents the number of samples in the pedestrian image data set, and the features in the feature set are All instances are stored in the feature storage unit, so that the category features in the feature storage unit can still be continuously updated when the clustered and unclustered outliers are constantly changing.

(2) Feature clustering and screening process

及未聚类离群值实例

其中n_c表示聚类的数量，n_o表示未聚类离群值的数量，随后依据聚类独立性与聚类紧密性标准，对聚类的结果进行筛选，并采用k-reciprocal近邻算法对检索结果进行重排序。First, the DBSCAN clustering algorithm is used to cluster the feature set {v ₁ ,...,v _n } in the feature storage unit in step 1, and the category prototypes in the feature storage unit are further divided into cluster centroids

and unclustered outlier instances

Among them, n _c represents the number of clusters, and n _o represents the number of unclustered outliers. Then, according to the criteria of cluster independence and cluster tightness, the results of the clustering are screened, and the k-reciprocal nearest neighbor algorithm is used to analyze the results. Search results are reordered.

In the DBSCAN algorithm, the maximum ε-neighborhood distance threshold is set to d=0.6, and the minimum number of neighbor points is set to 4, that is, a cluster contains at least 4 samples. By adjusting the ε-neighborhood distance threshold of the samples in the DBSCAN clustering algorithm, and alternately relaxing and tightening the clustering criteria, more reliable clustering results are obtained. Relax and tighten.

Both clustered and unclustered instances in the feature storage unit are regarded as equal and independent classes, so the reliability of clustering is critical to the impact of training, and the network has poor discrimination of images at the beginning of training. , the noise of clustering is also large, so a self-step learning strategy is proposed to improve the effect of clustering. Specifically, re-clustering is performed before the start of each round of training, starting from the most reliable clusters, retaining the reliable clusters, and disassembling the unreliable clusters back into the cluster-free outlier instances, gradually increasing number of clusters.

The cluster independence criterion is used to measure the distance between classes, which is expressed as the intersection ratio between the feature set and the feature set after relaxing the clustering criteria:

where |·| represents the number of features in the set, I(fi ) represents the set of samples in the same cluster, I _loose (fi ₎ represents the set of samples in the same cluster after relaxing the clustering criteria, R _indep (f _{i )} ) represents the independence score of cluster I(f _i );

The cluster tightness criterion is used to measure the intra-class distance, which is expressed as the intersection ratio between the feature set and the feature set after the clustering criterion is compressed:

where I _tight ( _fi ) represents the set of samples in the same cluster after tightening the clustering criteria, and R _comp ( _fi ) represents the tightness score of the cluster I ( _fi );

The independence between clusters and the closeness between samples can be measured through the above cluster reliability evaluation criteria. The starting point of the cluster reliability evaluation criteria is that a reliable cluster should be in a multi-scale clustering environment. It remains stable under the hyperparameter α,β∈[0,1] to denote the independence and closeness thresholds, where α is initialized as α=0.9*R _indep-1th , and remains consistent in subsequent training, R _indep-1th Represents the cluster independence criterion obtained in the first training round, β is set as the maximum cluster closeness criterion R _comp-max in the whole training process, in order to fully retain the most closely spaced samples in each cluster , retain the clustered samples for which the inter-class independence R _indep ( _fi )>α and the intra-class compactness R _comp ( _fi )>β, and the remaining samples are divided into unclustered outliers.

(3) Unified Contrast Loss Function Calculation Process

通过特征编码器编码后将特征全部存储到特征存储单元中，使自步学习策略将特征集合划分为聚类特征和未聚类离群值特征，从而整个训练数据集被分为有聚类伪标签的样本集合

和不属于任何聚类的离群值实例样本集合

且

使用特征存储单元用于存储样本特征的正面原型，对于一个来自无标注数据集的样本，若其在聚类内，则正面原型为其所对应的聚类质心，反之，若其不在聚类内，为聚类离群值，则正面原型为该离群值所对应的实例特征。Given unlabeled training samples

After encoding by the feature encoder, all the features are stored in the feature storage unit, so that the self-paced learning strategy divides the feature set into clustered features and unclustered outlier features, so that the entire training data set is divided into clustered pseudo A sample collection of labels

and a sample set of outlier instances that do not belong to any cluster

and

The feature storage unit is used to store the positive prototype of the sample feature. For a sample from an unlabeled data set, if it is in the cluster, the positive prototype is the corresponding cluster centroid, otherwise, if it is not in the cluster. , is a cluster outlier, and the positive prototype is the instance feature corresponding to the outlier.

使用特征编码器对每个训练样本进行前向计算得到编码的特征f，将特征f与其对应的类别原型进行向量点积运算计算相似度，构建统一对比损失函数：As shown in Figure 3, the schematic diagram of the unified comparison loss calculation, given the training sample

Use the feature encoder to perform forward calculation on each training sample to obtain the encoded feature f, and perform the vector dot product operation between the feature f and its corresponding category prototype to calculate the similarity, and construct a unified comparison loss function:

Among them, z ⁺ represents the positive category prototype of the feature f, τ represents the temperature coefficient, which is set to 0.05 according to experience, <·,·> represents the vector inner product, _ck is the centroid of the current cluster k, which represents the category prototype within the cluster , v _k is the instance feature of the current cluster k outliers, representing the class prototype without clustering;

If f belongs to cluster k, then z ⁺ =c _k is the centroid of cluster k; if f belongs to unclustered outliers, then z ⁺ =v _k is the instance feature of unclustered outliers, and the above contrast loss promotes coding After encoding the features of the mini-batch samples, they are compared with the prototypes of the two categories, so that each training sample is close to the category it belongs to and far away from other categories.

(4) Momentum update process of feature storage unit

As shown in the schematic diagram of the momentum update of the feature storage unit in Figure 4, firstly, the training samples are all stored in the instance unit, and then the features in each mini-batch sample are accumulated to the corresponding instance of the feature storage unit in the form of momentum update according to the index. Among the features, the momentum update method is widely used in the optimization algorithm in deep learning, which can keep the previous update direction to a certain extent when the parameters are updated, and at the same time use the characteristics of the samples in the current mini-batch to fine-tune the final update direction, in short It is to achieve the update of the current feature by accumulating the previous momentum.

而未聚类离群值的实例特征

直接从特征存储单元中提取剩余的实例特征，不失一般性，假定特征集合{v}中的未聚类离群值特征索引为{1,…,n₀}，那么聚类特征索引为{n_c+1,n}，第k个聚类的质心表示为：In the feature storage unit, the average value between the features is calculated for the features {v ₁ ,...,v _n } in the same cluster to obtain the cluster centroid

while the instance features of the unclustered outliers

Extract the remaining instance features directly from the feature storage unit, without loss of generality, assuming that the unclustered outlier feature index in the feature set {v} is {1,...,n ₀ }, then the clustering feature index is { n _c +1,n}, the centroid of the kth cluster is expressed as:

where I _k represents the feature vector set of the k-th cluster, |·| represents the number of feature vectors in the set, and the instance feature {v} in the feature storage unit is initially initialized once by the network forward calculation, and is continuously performed in the subsequent training process. Update for more robust clustering;

In each round of training, use the encoded features in the small batch of samples to update the category prototype in the feature storage unit in a momentum update manner, and perform a momentum weighted summation on the instance feature v _i and the current sample feature f _i to obtain the updated Instance features of :

v _i ←mv _i +(1-m)f _i

where m∈[0,1] is the momentum factor, which represents the proportion of instance feature _vi and sample feature _fi in the momentum update process, which is set to 0.2 according to experience. Given the updated _vi , if _fi belongs to If the cluster k is clustered, then the corresponding cluster centroid _ck is updated accordingly.

(5) Overall network training process

通过ResNet-50特征编码器进行特征编码，使用得到的特征集合初始化特征存储单元，使用Adam优化器进行模型参数的更新，权重衰减设为0.0005，训练轮回数设为70，采用学习率调度的方式进行模型的训练，以动态的方式来响应优化的进展情况，初始学习率设为0.00035，每20个训练轮回将学习率减小10倍。For unlabeled training datasets

Use the ResNet-50 feature encoder for feature encoding, use the obtained feature set to initialize the feature storage unit, use the Adam optimizer to update the model parameters, set the weight attenuation to 0.0005, set the number of training rounds to 70, and use the learning rate scheduling method The model is trained to respond to the progress of the optimization in a dynamic way, with the initial learning rate set to 0.00035 and decreasing the learning rate by a factor of 10 every 20 training epochs.

划分为聚类样本

和无聚类离群值样本

计算

中的聚类质心；In each subsequent training round, the feature set {v} in the feature storage unit is clustered according to the cluster independence and cluster tightness criteria, and the training samples are

Divide into cluster samples

and cluster-free outlier samples

calculate

the cluster centroids in ;

For training samples in each mini-batch, use the feature encoder f _θ for feature encoding, compute the contrastive loss, and perform backpropagation to update the encoder f _θ ;

Update the feature set v in the feature storage unit, and then update the cluster centroid _ck according to the updated feature set {v};

The feature encoder f _θ and feature storage unit are updated cyclically until the model achieves better convergence.

Finally, it should be noted that this embodiment provides the theoretical premise, implementation steps and parameter settings of an unsupervised pedestrian re-identification method based on contrastive clustering. In addition, this implementation is only a preferred specific implementation of the present invention, and in the specific implementation process, parameter settings need to be adjusted according to specific variables and data to achieve better practical effects.

Claims (7)

1. an unsupervised pedestrian re-identification method based on contrast clustering, is characterized in that, comprises the following steps: Step 1: Use the initial feature encoder to perform forward calculation on the unlabeled pedestrian image data set, and use the encoded features to initialize the feature storage unit based on the category prototype; Step 2: Before each round of training, perform clustering of the encoded features in the feature storage unit, and screen the clustering results according to the clustering reliability evaluation criteria; Step 3: Use the feature encoder to perform feature encoding on each group of mini-batch training samples, use the unified contrast loss function to perform back-propagation of the network, and update the feature encoder; Step 4: Dynamically update the feature storage unit in a momentum update manner using the encoded features; Step 5: Repeat steps 2 to 4 according to the number of training rounds until the pedestrian re-identification network converges. 2. a kind of unsupervised pedestrian re-identification method based on contrast clustering according to claim 1, is characterized in that, described step 1 comprises: Use a ResNet-50 deep neural network as a feature encoder and initialize it with pretrained weights on the ImageNet image dataset; Use the feature encoder to extract the features of the samples in the pedestrian image data set to obtain the feature set {v ₁ ,...,v _n }, and save all the sample features to the feature storage unit in the unit of instance. 3. a kind of unsupervised pedestrian re-identification method based on contrast clustering according to claim 1, is characterized in that, described step 2 comprises: Use the DBSCAN clustering algorithm to cluster the feature set {v ₁ ,...,v _n } in the feature storage unit in step 1, and the category prototypes in the feature storage unit are further divided into cluster centroids

and unclustered outlier instances

where n _c represents the number of clusters, n _o represents the number of unclustered outliers, using a self-paced learning strategy combined with cluster independence and cluster tightness criteria to retain reliable clusters, and Features in reliable clusters are disassembled back into cluster-free outlier instances. 4. The method according to claim 3, wherein the self-paced learning strategy re-clusters before the start of each round of training, starting from the most reliable cluster, and gradually increasing the number of clusters, through Adjust the sample neighborhood distance threshold in the DBSCAN clustering algorithm, and alternately relax and tighten the clustering criteria; The cluster independence criterion is used to measure the distance between classes, which is expressed as the intersection ratio between the feature set and the feature set after relaxing the clustering criteria:

where |·| represents the number of features in the set, I(fi ) represents the set of samples in the same cluster, I _loose (fi ₎ represents the set of samples in the same cluster after relaxing the clustering criteria, R _indep (f _{i )} ) represents the independence score of cluster I(f _i ); The cluster tightness criterion is used to measure the intra-class distance, which is expressed as the intersection ratio between the feature set and the feature set after the clustering criterion is compressed:

where I _tight ( _fi ) represents the set of samples in the same cluster after tightening the clustering criteria, and R _comp ( _fi ) represents the tightness score of the cluster I ( _fi ); Through the above cluster reliability evaluation criteria, the independence and closeness of the data in the cluster are measured, and α,β∈[0,1] is set to represent the threshold of independence and closeness, and the inter-class independence R _comp (f _i )>α and the intra-cluster tightness R _indep (f _i )>β of the clustered cluster samples, and the remaining samples are divided into unclustered outliers. 5. a kind of unsupervised pedestrian re-identification method based on contrast clustering according to claim 1, is characterized in that, described step 3 comprises: Given unlabeled training samples

Divide it into sample sets with clustered pseudo-labels using the self-paced learning strategy in step 2

and a sample set of outlier instances that do not belong to any cluster

and

Given training samples

Use the feature encoder to perform forward calculation to obtain the encoded feature f, and construct a unified contrastive loss function:

Among them, z ⁺ represents the positive category prototype of the feature f, τ represents the temperature coefficient, <·,·> represents the vector inner product, _ck is the cluster centroid of the current cluster, which represents the category prototype within the cluster, and v _k is the current cluster. Instance features of class outliers, representing class prototypes without clustering; After encoding the features of the mini-batch samples, they are compared with the two class prototypes, so that each training sample is close to the class it belongs to and far away from other classes. 6. a kind of unsupervised pedestrian re-identification method based on contrast clustering according to claim 1, is characterized in that, described step 4 comprises: In the feature storage unit, the average value between the features is calculated for the features {v ₁ ,...,v _n } in the same cluster to obtain the cluster centroid

while the instance features of the unclustered outliers

The remaining instance features are directly extracted from the feature storage unit, and the centroid of the kth cluster is expressed as:

where I _k represents the feature vector set of the k-th cluster, |·| represents the number of feature vectors in the set, and the instance feature {v} in the feature storage unit is initially initialized by the network forward calculation once, and is used in the subsequent training continuously updated in the process; All training samples are initialized to the feature storage unit in units of instances. In each subsequent round of training, the features in the current mini-batch sample are accumulated into the instance features corresponding to the feature storage unit according to the index, and the features in the mini-batch sample are used. The encoded features dynamically update the class prototypes in the feature storage unit in a momentum update manner: v _i ←mv _i +(1-m)f _i Where m∈[0,1] momentum factor, given the updated instance feature vi, if _{f i belongs} to cluster k, then the corresponding cluster centroid _ck needs to be updated accordingly. 7. a kind of unsupervised pedestrian re-identification method based on contrast clustering according to claim 1, is characterized in that, described step 5 comprises: In each training round, the features in the feature storage unit are clustered according to the cluster independence and cluster tightness criteria, and the training samples are

Divide samples into clusters

and cluster-free outlier samples

calculate

Cluster centroids in ; For the training samples in each mini-batch, use the feature encoder for feature encoding, calculate the unified contrast loss, and perform backpropagation to update the encoder; Update the feature set in the feature storage unit according to the category prototype momentum update method, and then update the cluster centroid in combination with the updated feature set; The feature encoder and feature storage unit are updated cyclically until the model achieves better convergence.

Images