Extraction And Fusion Of Diverse Features For Person Re-identification

Person Re-Identification(Re-ID)is a technology that uses computer vision technique to identify whether there is a target person in images or videos across different cameras.In recent years,the breakthrough of deep learning has promoted the rapid development of Re-ID,making it widely used in many fields such as criminal investigation,smart serveillance and security system,etc.However,complex environments such as illumination changes,occlusion and background transformations make the research on person Re-ID still very challenging.This thesis dedicated to improving the accuracy of person Re-ID by extraction and fusion of diverse features.The main contributions can be summarized as follows:(1)In order to achieve diverse features for person Re-ID,we propose a multi-branch network based on batch Drop Block.By varying dropping ratios for constructing diversity-achieving branches,feature diversity can be well achieved by data augmentation of batch Drop Block.Varieties of fusion schemes are compared and the integration of multiple global branches is proved to improve the accuracy significantly.Experiments show that the network we constructed can comprehensively improve the discrimination and generalization capabilities.(2)Compared with traditional networks such as ResNet,using Efficient Net is easier to obtain the optimal model through multi-factor optimization.In this paper,we study the influence of the depth,width,and resolution of the feature-extracting network based on Efficient Net,and propose a better backbone network for person Re-ID,termed Efficient Net-P.Experiments on multiple person Re-ID datasets show that using Efficient Net-P can achieve higher recognition accuracy without using of Image Net pre-training.(3)When multiple independently-labeled datasets are mixed together for cross-domain or multi-domain training,the domain gap makes the use of the current batch normalization module questionable.The distribution difference among training batches may result in unstable estimates for various statistical parameters,which may lead to possible performance deterioration for the employed batch normalization.Therefore,a novel domain-specific batch normalization algorithm is proposed for remedying the problem.Experimental results show that the proposed domain-specific batch normalization algorithm can effectively improve the final generalization ability under multi-domain training and obtain higher recognition accuracy.Furthermore,it can be applied to existing Re-ID networks for further improving performance.