| Monitoring can provide strong protection to people’s lives,especially in terms of security,such as the detection of criminal crimes,loss of the elderly,child abduction,etc.In the later stage,the identification is done by manual comparison or with the help of face recognition and other methods,but when encountering problems such as face obscuring,or the captured face image is not clear,the face detection system cannot play its role,and the manual one-by-one comparison is too labor-intensive.person rerecognition can complete the image retrieval task by using human body features,which provides the possibility to solve the above problems,so the research of person re-identification is of great significance.The development of person re-identification has made some progress so far,and has achieved good recognition results on various publicly available datasets,especially in the direction of supervised learning based on very high recognition accuracy.Supervised learning requires annotated datasets,and manual annotation of datasets is labor-intensive,so the research on person re-identification has gradually tilted toward unsupervised learning in recent years.In this paper,person re-identification with supervised learning and unsupervised learning are investigated respectively,and the main contents are as follows.In the supervised learning part,this paper proposes an improved model TransReID-KNN based on TransReID.since most of the current research on person re-identification is based on convolutional neural network as the main backbone network,but there are some inherent shortcomings of convolutional neural network,so this paper investigates in the supervised learning part the model TransReID-KNN developed based on Vision Transformer network.TransReID person re-identification model developed based on Vision Transformer network.In this paper,an improved model TransReID-KNN is proposed based on TransReID,and then comparison experiments are done on two different Vision Transformer backbone networks,through which the effect of TransReID-KNN model is verified and can effectively improve the accuracy of person re-identification.In the unsupervised learning part,this paper improves the unsupervised learning person reidentification based on patch from two aspects,one is to adjust the initial learning rate of the unsupervised learning person re-identification network based on patch,by increasing the initial learning rate can effectively improve the recognition accuracy and shorten the convergence speed significantly;the second is for the person re-identification dataset with low sample and subject to the GAN network enhancement dataset method.The effectiveness of the PAUL+CamStyle(PAULC)method on image level person reidentification is proposed,and the comparison experiments are done on the person re-identification original dataset,GAN generated dataset,GAN generated dataset fused with the original dataset and randomly removed part of the GAN generated dataset respectively.It is found that the GAN-generated dataset with reasonable deletion is used as the training set in this paper,which has obvious advantages in both recognition accuracy and convergence speed for unsupervised person re-identification based on patches.In this paper,improvement methods are proposed in both major research directions of person reidentification,and better performance is obtained with certain advantages. |
