| The person re-identification task aims to retrieve a given pedestrian image from a library of images captured by different cameras.However,in real scenarios,the pedestrian images captured by the cameras are usually of low resolution due to the quality of the camera equipment and the distance between the cameras and the pedestrians.And most of the image libraries are high-resolution images.When a low-resolution image is used to query a high-resolution image in the image library,it leads to the problem of resolution mismatch,which makes the accuracy of the person re-identification method decrease.In this paper,we conduct research on cross-resolution pedestrian re-recognition based on consistent representation learning,and analyze and design the algorithm in terms of taking into account three aspects: model performance,lightweighting and applicability.The specific methods are as follows.(1)Since the introduction of super-resolution methods to the cross-resolution person re-identification task,too much emphasis has been placed on training image super-resolution networks and person re-identification networks in cascade,which seems unreasonable.Moreover,in some cases,complex super-resolution networks do not necessarily perform better than simple super-resolution networks in the cross-resolution person re-identification task.Therefore,more thoughts are needed on the extraction of deep semantic and feature information of images.Based on the above analysis and reflection,a deep group-aware representation learning model for the cross-resolution person re-identification task is designed in this paper.The above possible problems are solved by using offline hyper-segmentation to generate super-resolution images by RCAN first,and then feeding the super-resolution images into the subsequent person re-identification network.Specifically,Transformer blocks with residuals are used in the feature extraction stage to extract richer local and global information from shallow images.This is the first attempt to apply Transformer to the cross-resolution person re-identification task.The multilayer group-aware representation case of the feature embedding module can generate a large number of representations different from the main branches,thus improving the quality of deeply embedded features.In addition,the kullback leibler scatter loss is designed and computed on the probabilistic prediction output of any two branches,allowing the whole network to be well optimized.Experiments on the MLR-Market-1501,MLR-CUHK03,MLR-VIPe R,and CAVIAR datasets validate the effectiveness of the method.(2)Many previous methods trained image super-resolution networks and person re-identification networks in cascade,which made the models very complex and unfavorable to the backward gradient propagation of the models,resulting in poor performance of the models.Moreover,the number of parameters and computation of these image super-resolution networks are relatively large,which is not favorable for deployment on some micro devices.To solve the above problems,this paper proposes an offline lightweight image super-resolution based cross-resolution pedestrian reidentification baseline network,which includes a well-designed lightweight image super-resolution multi-feature fusion network and a pedestrian reidentification baseline network AGW.on the MLR-Market-1501,MLR-CUHK03 and MLR-VIPe R datasets experiments validate the effectiveness of the method.(3)The offline hyper-segmentation method can effectively alleviate the problem of difficult backward gradient propagation due to cascade training.However,in real scenarios,sometimes the resolution of the query image is not known,and it is not known how to upsample it.Even if it is known,the low-resolution image is first oversampled and then recognized,which is a tedious operation process.Therefore,this paper proposes a multi-sensory cross-resolution person re-identification method based on intermediate domain,using an intermediate domain module,so that the low-resolution image can learn the lost detail information from the high-resolution image.And bridging loss and diversity loss are used to make the network generate suitable intermediate domains.Then,by using the group-aware learning network,the group-aware representation is integrated into the embedding features of the three domains,which greatly improves the recognition ability of deep embedding features.Experiments on the MLR-Market-1501 and MLR-CUHK03 datasets validate the effectiveness of the method. |
