Research On Unsupervised Domain Adaptation Person Re-identification Method Based On Generative Adversarial Network And Divergent Attention Mechanis

The goal of Person re-ID is match images of the same identity captured by different cameras of non-overlapping views.At present,the person re-ID method based on supervised learning has the greatest performance improvement in a single-domain,but when a model with high performance in a single-domain is directly applied to a new domain without labels,the performance will often decline.This is because samples from different domains have large domain gap due to different cameras,which is also a major reason why person re-ID cannot be effectively used in real life.To solve the above problems,academia has proposed unsupervised domain adaptive person re-ID methods,which can be divided into image transfer-based methods and pseudo labelbased methods.However,the existing unsupervised domain adaptive person re-ID methods have the problems that the domain gap is still large after image mapping and the feature extraction network cannot extract the fine-grained features of person,resulting in the generation of many pseudo labels with noise in clustering.To solve the shortcomings of the existing unsupervised domain adaptive person re-ID methods,starting from the above two methods to solve the unsupervised domain adaptive person re-ID,this paper proposes an unsupervised domain adaptive person re-ID method based on Generative Adversarial Networks and Divergent Attention Mechanism.The main work of this paper is as follows:(1)Aiming at the problem that the domain gap is still large after image mapping,this paper proposed an unsupervised domain adaptive person re-ID method based on generative adversarial networks.This method aims to map the sample style on the source domain to the style of the target domain and reduce the domain gap between the mapped images.In this method,the siamese convolution neural network is used to construct a person classifier to make full use of the identity label information.At the same time,the cycle generative adversarial network is used to transfer the image in the target domain to the image in the source domain,and the deep convolution generative adversarial network which can optimize the data distribution is added to optimize the domain mapping model to minimize the domain map between the mapped images.(2)Aiming at the problem that the existing unsupervised domain adaptive person re-ID methods based on attention mechanism and convolution neural network cannot extract the fine-grained features of person,resulting in many pseudo labels with noise in the process of clustering and generating pseudo labels,this paper proposed an unsupervised domain adaptive person re-ID method with divergent attention mechanism module and deep clustering module.This method can effectively encode fine-grained feature information,speed up the calculation time by reducing many matrix dot product operations,improve the accuracy of clustering and reduce the generation of pseudo labels with noise.