Cross Domain Person Re-identification Based On Generative Adversarial Network

Video surveillance has become an indispensable part of maintaining law and order and investigating cases in modern society,and it can accurately and effectively record audio and video information at a certain time.Through the analysis of surveillance information to screen,it is possible to complete a specific person search,route tracking,etc...However,the use of traditional methods or manual methods requires a lot of labor and time costs.The problem has been alleviated to some extent since the proposal of Person Re-identification(Person Re ID,Re ID specifically refers to person re-identification in this paper).However,compared with the current stage of research,the distribution of the feature space of the dataset has a large deficiency,and the dataset is difficult to produce,time and labor cost consumption and other characteristics,which makes the person re-identification in recent years in solving the difficult problem to stand still or slow progress.It is undeniable that the slow progress of person re-identification is related to the inadequacy of existing datasets for person re-identification.Generative Adversarial Networks(GAN in this paper denotes Generative Adversarial Network)although proposed after person re-identification.However,it has rapidly become another hot topic in computer vision and is widely used in many fields of computer vision,such as image restoration,image super-resolution,image retrieval,image generation,image editing,etc.How to use generative adversarial networks to perform data augmentation on existing person re-identification data,make up for the deficiencies of existing datasets through data augmentation methods,and finally achieve solutions to existing problems of person re-recognition,etc.In this paper,we will solve the deficiencies of the existing dataset through data augmentation in three directions: style,pose,and clothing.The main research contents of this paper are as follows.(1)To address the problem that the recognition rate of the model decreases when the illumination changes and the scene changes a lot,we use an improved current network for style transformation to perform stylistic data enhancement on the person reidentification dataset to obtain several times the amount of the original dataset under the target task.This task improves the Cycle GAN network,and we generate clearer images and feature spaces that are more consistent with the target feature space compared to the original network.By interconverting the data subsets in different feature spaces in the dataset,more training data can be obtained.At the same time,this data enhancement method can be used to transform the data into a realistic scene,which can be used to improve the recognition accuracy of the model in a certain scene,etc.(2)For the problem of fewer human poses in the dataset and more complex human poses in real scenes,the same operation as described in(1),we use the improved PATN to pass different images and their pose information through a specifically designed generative adversarial network to obtain images of a pedestrian image corresponding to other poses.This method can expand the data of human pose information in the dataset and finally achieve the effect of improving the accuracy of model recognition.(3)Similarly,for the problem of pedestrians changing clothes in real scenes.We can also generate images of one person wearing different clothes by virtual try-on.In this paper,we propose the PRQ-VTON network,which can generate virtual try-on effects with high quality and realism,which can be used to study pedestrian re-identification based on dressing change,and the method can effectively assist the study of pedestrian re-identification with dressing change.By applying different types of generative adversarial networks to the existing dataset,we can achieve data enhancement with different styles and poses on the existing dataset,and finally,complete the improvement of the accuracy of the training model.Besides,the virtual try-on technique in this paper can also be applied to the subsequent research on person re-identification based on the dressing.The relevant experiments on Market-1501,MSMT17,and other datasets in this paper can prove that the generated data quality is better than the current generated models,and the models trained with the expanded datasets have significantly improved in recognition accuracy.