Research On Unsupervised Cross-domain Person Re-identification Based On Deep Learning

Surveillance systems have spread across various public places,and video images taken have gradually become an important basis for preventing illegal crimes,criminal investigation tracking,and finding missing persons.When tracking person,person re-identification technology relies on its automatic retrieval to decrease artificial cost and speed up the retrieval process.However,the person re-identification model based on deep learning requires a large-scale dataset for training,and the production of the dataset is expensive.In addition,the various surveillance systems have differences in scenes,cameras,etc.,resulting in huge differences in styles between different datasets and cameras.These problems have seriously hindered the process of person re-identification technology apply in practical application.In order to reduce the impact of image style differences,this paper focuses on the unsupervised cross-domain person re-identification methods based on deep learning.Specifically,it uses the labeled data to train a model with excellent performance,and aims to obtain higher retrieval effect in unknown scenes.In the cross-domain learning,the cross-domain ability and cross-camera ability of the network are enhanced by domain adaptive loss.And this paper integrate the multi-level output of the network to comprehensively represent the person features.In the iterative training process,in order to decline the negative effect of the false label,this paper propose a dual network model of mutual supervision training to improve the accuracy of unsupervised cross-domain person re-identification.The main contributions of this article are as follows:(1)This paper proposes an unsupervised cross-domain person re-identification method based on three-stage training.It includes cross-dataset and cross-camera domain adaptation stage,self-supervised clustering training stage,representation loss and metric loss joint training stage.In the domain adaptation stage,cross-dataset adaptive loss and cross-camera adaptive loss are proposed to reduce the impact of domain differences.To obtain high-quality pseudo labels,this method performs self-supervised training on the basis of the pre-trained model,and enhances the domain transfer ability of the original model through joint loss training.At the same time,to reduce the impact of false labels in the model,the label smoothing regularization loss is used to replace the classical cross-entropy loss.(2)This paper proposes an unsupervised cross-domain person re-identification method based on multi-level feature clustering and mutual supervision training.To avoid incorrect clustering information being amplified iteratively,this method proposes a dual-network structure model for mutual supervised learning,which can effectively restrict the pseudo-labels generated by clustering to move closer to the real labels.In the pre-training stage,this method simultaneously performs the style transfer between the source domain and the target domain,which can effectively enhance the cross-dataset and cross-camera ability of the model.At the same time,the method uses the output of the multi-layer from network as the input of the clustering stage,which can comprehensively represent the person feature from multiple views,thereby obtaining high-quality clustering labels.