Research On Cross-domain Person Re-Identification With Complex Background

Person re-identification aims to match the images of the same person across non-overlapping cameras.It has broad application prospects in social management,emergency reconstruction and so on.With the development of deep learning technology,supervised person re-identification has achieved satisfactory results.However,in reality,building a labeled dataset for each new scene is very time-consuming and labor-consuming.Furthermore,when directly apply a model trained on a labeled domain to a new unlabeled domain,the performance of the model will decrease significantly.Aiming at the problem of weak cross-domain generalization ability faced by person re-identification models,the following researches are carried out in this paper:(1)Aiming at the problem of person image domain style difference caused by changes in lighting conditions,a person image light enhancement method based on Retinex convolutional neural network is proposed.Firstly,this method simulates the multi-scale Retinex algorithm and constructs a lightweight convolutional neural network for the mapping from original to enhanced image.Secondly,in order to train the light enhancement network without reference data,this method designs three non-reference losses for network training.Experiments show that this light enhancement method can effectively reduce the person image domain style difference caused by changes in lighting conditions.(2)Aiming at the problem of person feature domain gap caused by background noise,a Transformer based feature extraction network is proposed.This feature extraction network takes Transformer in the field of natural language processing as the basic framework and extracts global features,upper body features and lower body features of person images.Thanks to the multi head attention mechanism of Transformer,the network can explore the potential relationship between person local features and effectively suppress the background noise in features.Experiments show that the features extracted by the Transformer based feature extraction network are more discriminative and domain-invariant.(3)Aiming at the problem of pseudo label noise,a noise correction strategy and a noise restraint strategy are proposed.The noise correction strategy corrects the pseudo labels obtained by preliminary clustering based on neighborhood consistency,so as to reduce the noise in the pseudo labels.The noise suppression strategy gives different training weights to different samples by calculating the confidence of pseudo labels,so as to reduce the negative impact of pseudo label noise on network training.Experiments show that through these two noise refinement strategies,our person reidentification model can be effectively optimized with noisy pseudo labels.There are 24 figures,15 tables,and 80 citations in this thesis.