Research On Person Re-identification Algorithm Based On Deep Learning

Limited by the camera shooting height and the change of shooting perspective,the face features captured by surveillance are often not clear and complete,which greatly increases the difficulty of face recognition technology to be applied on the ground in real scenes.As an auxiliary supplement to face recognition technology,pedestrian rerecognition technology has received wide attention in computer vision field in recent years,aiming to solve the problem of recognizing and retrieving pedestrians under cross-camera and cross-scene,and has broad application prospects in such fields as intelligent security and cross-border tracking.Since the resolution varies from camera to camera and the captured pedestrian images are susceptible to problems such as occlusion,misalignment,background clutter,pose and dress variation,it has become important for pedestrian re-identification research to extract a more discriminative and robust feature.Recent studies have found that further mining the relationships between local features can provide more adequate semantic information for the final feature descriptors.Inspired by this,this paper proposes an end-to-end pedestrian re-recognition relational learning network to optimize the network model from two perspectives of local feature extraction and global feature extraction,and the specific design methods are summarized as follows:(1)For the local feature extraction branch,an efficient multi-granularity hypergraph relationship learning module is proposed to explore the interdependencies between local features.The module builds a hypergraph for each granularity,uses the hypergraph structure to adaptively select the neighbor nodes with the strongest relationship for each graph node,and conveys the higher-order relationships between nodes with the help of hyperedges.By mining the relationships between each graph node to model the internal structural information of the image,the discriminative nature of the feature vector can be effectively improved,which is conducive to accurately distinguishing different pedestrians with similar attributes.(2)For the global feature extraction branch,a hierarchical complementary recognition module is proposed,which consists of a main branch and an auxiliary line for extracting pedestrian salient features and sub-salient features,respectively.Guided by the main branch,the hierarchical complementary recognition module selectively activates different salient regions in the multi-scale feature map through saliency selection,saliency erasure and sub-salient mining operations,and uses residual learning to fuse the obtained salient features and sub-salient features to finally obtain a more comprehensive global feature.To verify the effectiveness of the designed relational learning network,this paper conducts sufficient experiments on three publicly available datasets,including the Market1501 dataset,CUHK03 dataset and Duke MTMC-re ID dataset.The experimental results show that the proposed method in this paper can significantly improve the performance of model retrieval,in which the m AP/Rank-1 evaluation metrics on the Market1501 dataset reaches 88.6%/95.4%,which is 3.2%/1.0% relative to Baseline,respectively.