Research On Cross-Resolution Person Re-Identification Based On Deep Learning

In live-action person re-identification(re-id)applications,it is common to use low-resolution(LR)query images to match high-resolution(HR)gallery images,which will lead to the resolution mismatch between query and gallery images.The accuracy of person re-identification(Re-ID)is subject to the limitation of camera hardware conditions and the change of image resolution caused by factors such as camera focusing errors.People call this problem cross-resolution person Re-ID.The existing methods mostly use the resolution of the image super-resolution(SR)technology to restore LR images to keep the query image and gallery image of the same or similar highresolution to perform re-id tasks.However,it is unreliable to rely on the HR image generated by the generator to perform re-id tasks,which is also the limitations of most research at present.This paper proposes two solutions to the problem of cross-resolution person re-id:(1)We treat cross-resolution person Re-ID as a two-stage task: the first stage is the image enhancement stage,and we propose a Super-Resolution Dual-Stream Feature Fusion sub-network,named SR-DSFF,which utilizes the SR module restores the resolution of the LR image,and then obtains the features of the low-resolution and HR images respectively through the two attentionweighted feature extraction streams in the DSFF module,and fuses the feature maps.Set the convolution to restore the final images.The second stage is the feature acquisition stage.We design a network for a global-local feature extraction network guided by human pose estimation,named FENet-Re ID,FENet-Re ID obtains the final feature maps through multi-stage feature extraction and multi-scale feature fusion for the Re-ID task.(2)We propose a Triple-stream Joint Multi-resolution Pyramidal-feature Learning(TJMPL)Framework.Initially,we add the Coordinate Attention(CA)mechanism to the original MDSR module and named the new SR module MDSR-CA.This MDSR-CA module is more adaptable for the specific re-id task while restoring the resolution of LR images.Next,we jointly learn the specific features from the LR and SR image and fuse the features into a discriminative feature.This joint learning process is mainly implemented through a Low-resolution Pyramidal-feature Extraction(LPE)stream,a Super-resolution Pyramidal-feature Extraction(SPE)stream,and a Joint Learning Feature Fusion(JLFF)stream in TJMPL.Subsequently,we propose a Feature Fusion Unit with Attention(FFUA)module in the JLFF stream to better fuse the specific features from different resolution images.Furthermore,we explore the role of multi-scale feature information in joint multi-resolution learning by slicing and recombining a feature map into a pyramidal 3-dimensional sub-map group.We conduct extensive experiments on multiple datasets to verify the advancement of the proposed methods.A large number of experimental results show that compared with the existing advanced methods,the proposed method can better solve the problems encountered in crossresolution person re-identification and achieve more advanced performance.