Research On Dual-path Visible-Infrared Person Re-identification Based On Generative Adversarial Network

With the deepening of research,the field of person re-identification has developed rapidly,and this technology has also been widely used.When the light conditions are good,the visible light person re-identification method performs well even in the face of complex environment.However,the problem of insufficient illumination often exists in practical application,the resolution of visible light images taken by common ordinary cameras is extremely low,which can negatively affect the person re-identification task,so it is necessary to use infrared cameras to shoot additional clear infrared images.Due to the significant difference between the imaging spectrum of infrared image and visible image,there is a huge modal difference between them.In order to solve the above problems,this thesis proposes an image pair generation network and a dual-path joint discrimination model based on the common research methods and related concepts of person re-identification research.The main innovations are as follows:(1)The thesis proposes an image pair generation network(IPGN)to solve the problem of modal inconsistency in visible-infrared person re-identification.The network uses attribute encoder and modal encoder respectively to divide the image into two parts,attribute code and modal code respectively.For visible and infrared images with the same person identity,the modal codes of different images are exchanged and input into the generator together with the original attribute codes of the images,an image corresponding to another modality therefore is generated for each modality.The images with different attribute but the same modality form the same modality image pair,so the problem of feature misalignment caused by different modalities is solved.Experiments are designed on Reg DB and SYSU-MM01,and the results prove the effectiveness of IPGN method.(2)The thesis proposes a dual-path image pair joint discriminant model(DPJD).The model includes two discriminant branches,one with the same modality and the other with different modalities.In the discriminant branch with the same modality,the model mainly focuses on the attribute information of images.In the discriminant branch with different modalities,the model mainly focuses on modal information of images.The two branches act on the discriminant module to make full use of all the features of the image.Since DPJD processes both modal information and attribute information,the model uses KL divergence to enhance the correlation between different features.In addition,the model uses triplet loss to optimize the intra-class distance and increase the inter-class distance,so as to improve the overall discrimination accuracy.Experiments are designed on two kinds of benchmark data sets and verify the superiority of DPJD method.