Research On Unsupervised Person Re-Identification Technology Based On Feature Refinement

Person re-identification has attracted much attention as a core technology in the fields of intelligent monitoring,public security criminal investigation,and regional security.In recent years,the development of deep learning has made supervised person re-identification methods obtain higher recognition accuracy.However,in practical applications,supervised person re-identification methods use manual labeling of camera views at a high cost,and the network performance is highly dependent on the labeled samples in the data set,and the scalability is not high.The unsupervised person re-identification method does not require artificially labeled identity information,which reduces the heavy workload of manual data annotation and has more development potential in large-scale actual scenarios.However,the existing unsupervised methods are difficult to obtain superior performance compared with supervised learning.On the one hand,in the absence of reliable annotation information,semantic information has not been further refined and fully utilized.On the other hand,various appearance changes caused by pose changes and local differences make it difficult to correctly identify certain types of person.This paper mainly focuses on the refinement features between unlabeled images.Therefore,an unsupervised person re-identification technology based on feature refinement is proposed.The main research contents and work are as follows:(1)Most of the traditional unsupervised person re-identification methods focus on the reliability of person pseudo-labels,and it is easy to ignore the negative impact of refined feature differences on the performance of unsupervised classification networks.From the perspective of global features,an improved unsupervised person re-identification method based on refined features guided multi-label distribution ranking learning is proposed.This method alleviates the problem that MMCL algorithm pays insufficient attention to the potential key areas in person images and the recognition difficulty of positive and negative samples is not balanced.Firstly,a refined feature extraction module is designed and embedded into the original ResNet50 network to obtain refined feature information.Secondly,constraints are imposed on the ranking between positive and negative category labels in multi-category labels,and reliable multi-category pseudo labels are selected.Finally,the multi-label distribution ranking loss and multi-label classification loss are combined to supervise the network and improve the overall performance of the model.The proposed method achieves 58.4%and 42.3%recognition accuracy in the Market-1501 and DukeMTMC-reID datasets,respectively,and the average accuracy on the Market-1501 dataset is improved by 12.9%.(2)With the development of unsupervised person re-identification technology,the method of generating pseudo-labels by clustering has achieved considerable performance.However,due to the insufficient feature expression ability of the unsupervised model and the insufficient semantic information of the features extracted from the samples,it is easy to cause noise pseudo-labels.To this end,an improved unsupervised person re-identification method combining feature refinement and noiseresistant comparative learning is proposed.This method mainly captures the discriminative representation of unlabeled data by fusing the important features of non-local and channel,and weighted strengthens the refined feature information to form a more discriminative feature descriptor.In addition,by designing a noise-resistant dynamic contrast equalization loss function for unsupervised joint learning,the negative impact of noise labels on the network is reduced.The experimental results on Market-1501 and DukeMTMC-reID datasets verify the effectiveness and advancement of the proposed method.The average accuracy mAP reaches 83.1%and 71.9%respectively,and the first hit rate Rank-1 is 93.5%and 85.6%respectively.(3)In real intelligent surveillance scenarios,unsupervised person reidentification cannot obtain sufficient feature expression from a single unlabeled global image due to illumination differences,occlusion and person posture changes in the environment.When the sub-significant areas in the image are ignored,it is easy to weaken the key information and affect the final performance of the model.Aiming at the above problems,an improved unsupervised person re-identification method based on local refinement multi-branch and global feature sharing is designed.This method no longer follows the single feature description in traditional unsupervised learning,but combines rough global features and fine features in local refinement branches to obtain person diversified feature expression.In addition,an attention block of channel refinement information fusion is designed before the branch operation to enhance the network ’s attention to person features and perform focused learning of refined features.The average accuracy of the proposed method on Market1501,DukeMTMC-reID and MSMT17 datasets is improved by 4.4%,3.2%and 6.4%respectively,and the average accuracy on Market-1501 dataset is 83.3%.