Research On Cross-Domain Person Re-Identification Based On Generative Adversarial Network

In real life,the public security departments usually search for the traces of criminal suspects through video detection and video analysis.In the current era of artificial intelligence,this method of relying solely on manual detection is obviously not suitable for the trend of technology.It is more time-consuming and laborious,and easy to cause visual fatigue of the staff.It is very likely that the best time to capture is missed.How to use technology to achieve fast retrieval has become a problem.Person Re-identification(Re ID)has received more and more attention in security fields such as intelligent surveillance and video tracking because its goal is to retrieve all pedestrians similar to a specific pedestrian,which can better assist security personnel in monitoring and searching.However,in actual scenarios,the research on Re ID is still faced with many challenges,such as: a Re ID model that performs well on one dataset is tested on another dataset(cross-domain).Insufficient richness,blurred images,cluttered backgrounds,and pedestrians occluded by foreign objects.Therefore,if the Re ID technology can be widely used in real life like the face recognition system,the most important task is to solve these problems.This paper uses Generative Adversarial Networks(GAN)to solve the problems of insufficient data,blurred images,pedestrians occluded by foreign objects,and insufficient pedestrian features in a cross-domain background.A Cross-domain Person Re-identification Model Based on Generative Adversarial Network(CDGAN)is proposed.The main research contents are as follows:(1)Aiming at the problem of insufficient data,this paper proposes a Person Re-identification Network Based on Pose Transfer(PTN).We use GAN to synthesize pedestrian images and different target poses to generate images with different poses,which not only enriches the poses of pedestrians,but also expands the dataset,so that the model learns more samples and is more recognizable.(2)Since all the images in the original dataset have blur problems,this leads to the blur problems in some of the images generated by the above-mentioned PTN network.This paper proposes a Person Re-identification Network Based on Super Resolution(SRN).Using GAN to reconstruct the resolution of the image makes the image clearer and the network is more discriminative when learning image features.(3)Aiming at the problem of insufficient image features extracted by the network,this paper proposes a Cross-domain Person Re-identification Network Based on Feature Fusion(FFGAN).The network separately extracts the global features,local features and semantic features of pedestrians,and fuses the three types of features into a complete expression of pedestrian features.Before extracting features,in order to prevent the image from being affected by occlusion factors,this paper proposes a feature erasing module to directly erase this part of the occluded area.At the same time,in order to obtain more key features of the network,this paper uses the complementarity of hard attention,soft attention,spatial attention and channel attention mechanisms to combine into a unified attention mechanism module.Experiments prove that FFGAN has complete expression ability and better fitting features.Finally,this article puts the dataset generated after PTN enhanced data and SRN to improve image resolution processing into FFGAN for training.After many experiments,the three networks have effectively improved the accuracy of the model.It is in the context of solving cross-domain problems.At the same time,it solves the problems of incomplete model extraction features,insufficient training samples,blurred images,and occluded images,and improves the robustness and generalization of the cross-domain model.