Research On Person Re-identification Based On Generative Adversarial Network

Person re-identification(re-id)aims to establish the identity association of persons in cross-camera scenes,which retrieves images of a target person from a large gallery of images captured by non-overlapping cameras based on a specified query image.Person reidentification has important applications in fields such as intelligent security,but the problem of insufficient data and low data diversity in training datasets leads to poor identification performance of existing methods in practical scenarios.In particular,the large differences in background,perspective,and pose of images captured by different cameras further increase the difficulty of identifying target persons in cross-camera scenes.Therefore,it is of great theoretical significance and application value how to improve the re-id performance of person re-identification models under data lacking scenes.To address the issues that previous approaches based on generative adversarial networks(GANs)cannot properly guide the synthesis of identity features,and that the separation of data generation and re-id training limits the performance of person re-id,this work offers a person re-id model(IDGAN)based on semantic map guided identity transfer GAN.With the aid of the semantic map,IDGAN generates pedestrian images with varying poses,perspectives,and backgrounds efficiently and accurately,improving the diversity of training data.Then,IDGAN adopts a gradient augmentation method based on local quality attention,which feeds richer gradient information for local regions with poorer quality by evaluating the generated images,so that the generator can focus on both global and local regions and improve the visual realism.To make full use of the generated data,IDGAN employs a two-stage joint training framework to allow the generative adversarial network and the person re-id network to learn from each other.Detailed experimental results demonstrate that IDGAN is capable of producing highquality images and significantly enhancing re-id performance,with the FID(Fréchet Inception Distance)of generated images on the Market-1501 dataset being reduced by 1.15,and m AP(mean Average Precision)on the Market-1501 and Duke MTMC-re ID datasets being increased by 3.3% and 2.6%,respectively,in comparison to existing methods.In this paper,a person re-identification system for a real-world scenario is constructed,with a detailed requirements analysis,overall design,module implementation,and performance testing.The system implements and integrates the proposed person reidentification model,target detection model,and target tracking model.The system test results verify the accuracy and effectiveness of the proposed person re-id model IDGAN.