Person Re-identification Based On Joint Learning Of Generation And Discrimination

Person re-identification is to use computer vision technology to classify the person images captured by specific cameras in the scene captured by multiple cameras,and retrieve the person image pairs whose identities are consistent in the large-scale person image database.It is widely used in intelligent security,smart city and other fields.In the task of person re-identification,there are some problems,such as the difficulty of datasets annotation,the small sample size and so on.The image generation and discrimination are two independent stages,which can not make full use of the generated image,and the detail feature extraction is insufficient.To solve the above problems,this paper proposes person re-identification based on joint learning of generation and discrimination.Firstly,the framework of teacher-student joint learning is constructed.The image discrimination module is directly embedded in the generation module by sharing the appearance encoder with the generation module.In this way,the image generation module and discrimination module are integrated into a framework to realize the end-to-end training of image generation and discrimination.In the process of training,the generated image is fed back to the appearance encoder in real time,which can not only optimize the appearance feature vector generated by the appearance encoder,but also optimize the recognition accuracy of the discrimination module.Image generation includes three methods: first,the image is generated by exchanging the appearance and structure features of the same image with consistent identity,so as to regularize the generator;Secondly,the appearance and structure features of the two images are exchanged to ensure that there is no problem in the original image generated by the generation module;Thirdly,two images with different identities and different images are exchanged to synthesize a new image to ensure that the generated image is controllable and matches the real data,and the synthesized image is consistent with the real image identity information that provides the appearance structure.Through the above three image generation methods,the sample size of the data set is increased,which makes up for the problem that the training set is difficult to recognize due to the similar appearance.Secondly,a deep dual attention mechanism(DDAM)is proposed,which combines attention connection network with dual attention mechanism(DAM).The convoluted feature map is used as the input of the channel attention module,and the feature map with channel attention is obtained through the channel attention module.The feature map with two-dimensional spatial attention is obtained by taking the feature map as the input of the spatial attention module.Then,the attention modules of adjacent channels and adjacent space are connected to each other,so that the information of attention modules of the same dimension can be exchanged.All attention modules are trained together to ensure the flow of information in the way of feed forward,avoid the problem of frequent changes in the process of information transmission,and improve the learning ability of attention modules,In order to get better attention characteristics.Finally,the deep dual attention mechanism is embedded into the teacher model to further improve the ability of the teacher model to assist the student model to learn the main identity information of pedestrian image,and then combined with the fine-grained feature information learned by the student model,which has nothing to do with the appearance features,to improve the model recognition ability.The experimental results show that the image generated by the proposed model has diversity,and avoids the interference of noise in image generation,and the extracted features have strong robustness.The experimental results on the Market-1501 and DukeMTMC-ReID datasets show that the model has good discrimination at the same time.