Promoting Convolution Neural Network Based High Precision Person Re-identification

In recent years,with the development of national technology and economy,surveillance cameras have been widely deployed in hospitals,schools,airports and other public places.However,the rapid increase in the number of surveillance cameras has also brought massive amounts of surveillance data,which has brought great challenges to traditional manual processing.Therefore,intelligent security systems have emerged.Person re-identification technology is an important part of intelligent security systems.It uses computer vision algorithms to determine whether there are specific pedestrians in surveillance photos or surveillance videos.Because of its huge research and practical value,it has attracted a lot of attention from academia and industry,and it is currently one of the important research topics in the field of computer vision.However,in the realistic application deployment of the pedestrian re-recognition model,it will still be affected by three complex factors,i.e.,crosscamera,cross-domain,and cross-resolution,resulting in low recognition accuracy.This article has conducted in-depth and systematic research on these three issues,aiming to improve the accuracy of person re-identification based on deep learning.The main research contents and innovations are as follows:1.Background Gradually Suppression and Feature Pyramid Optimization based CrossCamera Person Re-Identification.Person re-identification usually faces internal and external challenges.The internal challenge is that the pedestrian background constantly changes in the realistic scenarios across cameras,and these background noises have a serious impact on the training and application of the model.In response to this problem,this paper proposes a background gradually suppression strategy.Person images are splitted using target detection algorithms and semantic segmentation algorithms into foreground and background,the background partition are assigned with different weights and then are input into different convolutional neural network branch for training.The external challenge is the lack of discrimination of the features when the model is deployed in the realistic scenarios.In response to this problem,this paper proposes a feature pyramid optimization strategy,which aims to solve the error back propagation problem of traditional feature pyramid without increasing network parameters.Extensive experiments have been conducted and demonstrate the two strategies proposed in this paper can effectively improve the accuracy of person re-identification in cross-camera scenes.2.Style Transferring based Cross-Domain Unsupervised Person Re-Identification.Aiming at the problem that large-scale person data cannot be annotated in realistic scenarios,this method uses a Cycle GAN to transfer the labeled database and the unlabeled database,and uses the labeled database after the style transferring to train the model.To transform the cross-domain unsupervised person re-identification problem into a supervised problem.Since the process of style conversion is unsupervised,it will lead to the image of the same pedestrian to be transferred with the image of different pedestrians,resulting in the destruction of the identity feature of the pedestrian.In response to this problem,this paper embeds a self-labeled triplet network to the Cycle GAN to protect pedestrian identity feature.At the same time,this method introduces the maximum mean discrepancy loss.During the training process,the model will narrow the distribution of the unlabeled dataset and the labeled dataset.Extensive experiments have been conducted and demonstrate both the self-labeled triplet network and the maximum mean discrepancy loss can improve the accuracy of crossdomain unsupervised person re-identification.3.Distribution Unifying based Cross-Resolution Person Re-identification.Affected by complex factors such as light,distance,and camera pixels,the resolution of pedestrian images obtained in realistic scenarios usually varies greatly.Conventional methods to solve the cross-resolution person re-identification problem usually use a super-resolution module to increase the resolution of low-resolution images,and then use the person re-identification module to train the super-resolution images to obtain a model.However,these methods neglect the extensive existing distribution differences in cross-resolution person re-identification tasks.In response to this problem,this paper proposes SRMMD(Super Resolution Maximum Mean Discrepancy)loss and Re IDMMD(re-identification Maximum Mean Discrepancy)loss to two modules,respectively.In the super-resolution module,the super-resolution images and the high-resolution images are used to compute the maximum mean discrepancy loss,to narrow the distribution distance of super-resolution images and high-resolution images from the image level.In the person re-identification module,two convolutional neural network branches are used to train high-resolution images and super-resolution images,respectively.The maximum mean discrepancy of the features obtained from the two branches is calculated to narrow the distribution distance at the feature level.Extensive experiments have been conducted and demonstrate both SRMMD loss and Re IDMMD loss can effectively improve the accuracy of person re-identification in cross-resolution scenes.