Research On Person Re-identification Algorithms Based On Convolutional Neural Network

In recent years,with the rapid development of deep learning and computer hardware,person re-identification plays an important role in intelligent monitoring and intelligent security,attracting extensive attention from researchers.Compared with the previous methods based on hand-crafted features,the person re-id methods based on CNN,mining information from largescale person image data,has better feature representation capability.Many recent works achieve good results in this field,but there are still some problems need to be solved.In the research of person re-id,according to the level of supervision,it can be roughly divided into supervised person re-id,unsupervised domain adaption person re-id and unsupervised person re-id.In order to solve the main problems in these three kinds of person re-id tasks under different levels of supervision,based on the current related works,this paper proposes three new methods for these three kinds of person re-identification tasks under different levels of supervision:(1)For the supervised person re-identification task,the Mixed-Domain Relation Attention Network is proposed,which can effectively suppress complex but useless background information and alleviate the influence of obvious intra-class variance through the MixedDomain Relation Attention(MRA)module.The MRA explicitly exploits global relational patterns to mine structural information,which is helpful for inferring discriminative high-level semantic features.Benefiting from the mining of global relational patterns,the model’s attention can be focused on key human body regions.In this attention mechanism,for each feature node,the pairwise relationship between the node and all other nodes is modeled,and the relationship features are compactly stacked into tensors,and then combined with the features of the node itself,through a small model to infer attention weights.In this way,the MRA comprehensively utilizing local and global information,can better get rid of the interference of useless information,and finally weight the importance of feature nodes from a global perspective.(2)For the unsupervised domain adaption person re-identification task,a cluster-level contrastive learning framework is proposed to optimally learn noise-tolerant representations on unlabeled target data.It is mainly aimed at two limitations of existing research: one is that the label noise introduced by generating pseudo labels on the target domain hinders model optimization;the other is the obstacle to knowledge transfer caused by the gap between the source domain and the target domain.Aiming at these two problems,this paper proposes and integrates three methods.First,a cluster-level contrastive learning method is proposed to learn noise-tolerant representations in an unsupervised manner by iterative optimization of feature learning and cluster refinement.Second,a progressive transfer strategy is proposed,which not directly fine-tuning the model in the training stage,but adopts a collaborative learning mechanism of shared feature encoders on both domains.By gradually reducing the training weight on the source domain while increasing the training weight on the target domain,knowledge transfer from the source domain to the target domain can be better achieved.Finally,a Fourier-augmentation method is proposed to maximize the class separability by adding extra constraints in the Fourier space.Combining these methods,the model can learn more discriminative unsupervised noise-tolerant representations on the target domain.(3)For unsupervised person re-identification task,an unsupervised person re-identification algorithm based on soft labels learning is proposed,which alleviates the adverse effect of label noise introduced by pseudo labels on model training.By mining the relationship between the feature vectors of unlabeled pictures in the mapping space,similar pictures are made closer in a soft constraint manner,and the feature encoder network is trained with soft labels reflecting the similarity of the images.Unlike the one-hot labels that force images to belong to a specific class,the soft labels used can be viewed as a probability distribution on multiple classes.The training goal of the network is not only to predict the ground truth,but also to predict the similar classes.And the final obtained feature maps of similar images are closer to unrelated images and have a larger distance.In addition,some auxiliary information is also introduced to help find similar images.The id of cameras and local details of pedestrian images are also utilized when measuring the similarity between images.A cross-camera encouragement mechanism is proposed to alleviate cross-camera differences in images of the same person.Combined with global appearance,local details and cross-camera encouragement,the model shows good performance on several datasets.In summary,this paper proposes three person re-identification methods based on CNN.Through experimental verification on the public datasets Market 1501,CUHK03,Duke MTMC-re ID,MSMT17,the proposed methods achieves better results compared with many current methods and has good application value.