Unsupervised Person Re-identification Based On Deep Learning

Person re-identification(person Re-ID)is the task of retrieving specific pedestrians in non-overlapping camera scenes,which has great research significance in intelligent security and other fields.Compared with supervised person Re-ID,unsupervised person Re-ID can use dataset samples without labeling,which is more suitable for real-world applications.Due to the influence of style difference between datasets,and the noise of false labels generated by clustering and other methods,unsupervised person Re-ID faces great challenges.This paper mainly studies the unsupervised domain adaptive person ReID and proposes solutions to the above problems.The main work is summarized as follows:(1)To solve the problem of style difference between datasets,BN layer and IN layer are connected to form IBN network module,so as to form IBN-Res Net-50 network,which serves as the backbone network for feature extraction of person Re-ID model.The results of supervised source domain experiments show that IBN-Res Net-50 learns more features than Res Net-50,which is more conducive to classification.The results of unsupervised domain adaptive experiments show that the IBN-Res Net-50 model is less affected by the style difference between datasets,and improves the performance of re-identification in the target domain.(2)To solve the problem of noisy pseudo-labels generated by clustering,this paper proposes an unsupervised person Re-ID model based on Label Denoising(LD).The neighbor label correction module is designed based on the similarity ranking of sample features,which allocates different numbers of nearest neighbor samples to target samples by referring to the in-class numbers.The pseudo label of the target sample is corrected by nonlinear marking system of the nearest neighbor samples.The center shift loss function is designed to make the cluster more compact by shortening the distance of same label samples.The results of experiment show that the LD model improves the accuracy of pseudo-labels in the target domain and has a certain ability of label denoising.(3)In order to make the model have adaptive anti-noise capability,this paper proposes an unsupervised person Re-ID model based on Dual Network Cooperative Training(DNCT).The model builds the mean teacher network and updates the parameters of the teacher network by EMA of the student network.According to the consistency prediction of the same sample by dual networks,the uncertainty calculation module is designed to calculate the reliability of the sample and weight it on the hard label loss function.In order to strengthen the connection between the dual networks,the soft label loss function is designed to supervise the student network training by using the prediction information of teacher network.The results of experiment show that DNCT model can recognize noise samples and improve the performance of re-identification.