The Person Re-identification Algorithm Based On Unsupervised Learning

Person Re-Identification(Re-ID)aims to identify the designated query person from a large number of candidates captured by non-overlapping cameras,which is widely used in many fields such as intelligent security and intelligent supermarket.With the development of deep learning technology,supervised person Re-ID has achieved impressive performance,but the heavy label burden makes it difficult to apply to practical problems.Therefore,in recent years,many researchers began to study unsupervised person Re-ID.Unsupervised person Re-ID includes unsupervised domain adaptive(cross-domain)person Re-ID and pure unsupervised person Re-ID.One of the key problems in unsupervised person Re-ID is how to minimize the influence of noise pseudo labels.To address this problem,this paper conducts research from two aspects:improving the accuracy of pseudo labels and directly suppressing the influence of noise pseudo labels.The three works of this paper are as follows:Aiming at the problem of limited receptive field and lack of interaction between local information modeling and global information modeling,this paper proposes an interactive cascade microformers.As the backbone network,the proposed microformer model can effectively capture the long-distance dependence,as well as avoids the aggravation of the influence of noise labels.The proposed interactive cascade framework can make use of the interaction between local information modeling and global information modeling by alternately performing these two modeling process.The interactive cascade microformers can effectively extract the recognizable person features,thus improving the accuracy of pseudo labels.Aiming at the problems of low illumination and distortion that limit the accuracy of pseudo labels,this paper introduces image quality enhancement algorithm into unsupervised learning person Re-ID task for the first time,which reduces the difficulty of extracting features from person images and improves the accuracy of pseudo labels.Specifically,in this work,low-light enhancement algorithm and detail information enhancement algorithm are used to generate quality enhance images,and then these two kinds of images are superimposed as the input of the person Re-ID backbone model,so that the backbone model can effectively extract person features,and this scheme can also generate more accurate pseudo labels.Aiming at the problem that noise pseudo labels affect the model learning process,this paper proposes a series of adaptive mechanism.Specifically,the relative distances between positive hard samples and class prototypes are employed as an index to assess the credibility of positive hard samples,and the updated degree of class prototypes is adaptively changed based on this index.In addition,the model’s attention to the hard samples is adaptively changed based on the distance between the positive hard samples and the class prototypes.We conduct systematic experiments on the benchmark datasets.The experimental results show that the methods proposed in this paper have achieved better performance,which verifies the effectiveness of the proposed methods in unsupervised person Re-ID task.