Research On Person Re-identification Based On Self-supervised Learning

Person re-identification is an important research topic in the field of computer vision and intelligent security,which aims at identifying the individual of interest in a set of images across different cameras.In recent years,its performance has been greatly improved due to the development of deep learning technology.However,due to pose changes,illumination variations,background and occlusion interference,person re-identification research is still far from practical applications.Self-supervised learning is an emerging branch in the field of deep learning.It holds better scalability when processing large-scale data for the ability of mining potential supervision from the data itself.Based on self-supervised learning,this dissertation proposes corresponding solutions to some specific problems.The main work and innovations of this dissertation include the following four parts:(1)For the problem of insufficient training data,a four-branch Siamese network-based architecture is proposed.Firstly,we design a data augmentation strategy based on self-supervised learning,which is achieved by horizontally dividing pedestrian images into upper and lower parts,and recombining the upper and lower bodies of different pedestrian images.Two specific data augmentation methods can be achieved based on this process,i.e.,offline ID augmentation and online instance augmentation.Specifically,ID augmentation can generate new IDs,and instance augmentation can produce new images based on the existing IDs.Moreover,combining with the proposed data augmentation,a suitable four-branch Siamese network is designed and integrated with the above two data augmentation components in a unified framework to improve the performance of person re-identification.(2)For the problem of how to extract diversified semantic features,a self-supervised based feature disentangling method is proposed.Firstly,we achieved a simple and efficient data augmentation method based on the correlation between the channel information of the three-channel color pedestrian image itself,which can generate a large number of training samples sharing similar edge information.Meanwhile,a soft label assignment strategy is designed to characterize the correlation between the original samples and the corresponding generated ones.Then,according to the proposed data augmentation,this dissertation designs a network based on the encoder-decoder structure and achieves the separation of different semantic information through feature disentangling based on components exchange.The data augmentation in this method can provide data support for subsequent feature disentangling.Therefore,the proposed feature disentangling method is more interpretable and can effectively improve the performance of person re-identification.(3)For the problem of how to utilize the idea of transfer learning to improve the performance of supervised person re-identification,a self-supervised based negative transfer learning method is proposed.Firstly,the method of image channel shuffling is used to generate multiple augmented counterparts from the source domain.This operation provides a data basis for subsequent transfer learning.Then,in view of the symmetry between the source domain and the augmented domains,a symmetric classification network for person re-identification is elaborated,and negative transfer learning is applied to train the network based on the principle of structural consistency across domains.Different from traditional transfer learning methods that transfer knowledge from the source domain to the target one,our method promotes the performance in the source domain by urging the model to converge in the source domain and the augmented domain simultaneously.Therefore,it is called negative transfer learning.(4)For the problem of how to mine potential supervision information from the data itself in unsupervised person re-identification,a self-supervised multi-domain joint learning method is introduced.Firstly,data augmentation is achieved in the pre-training stage,and multiple label assignment strategies are designed according to the characteristics of augmented data to focus on different aspects of the data,and then feature decoupling is achieved through multi-task learning.Furthermore,in the stage of model fine-tuning,according to the logical relationship between different features,a variety of clustering algorithms are selected to cluster different semantic features separately.Finally,the correlation between different clustering results is used to further distill the clustering results to improve the reliability of the pseudo labels,and ultimately improve the performance of unsupervised person re-identification.