Research On Person Re-identification Based On Associative Feature Fusion And Multi-scale Learning

Person re-identification technology aims to solve the problem of retrieving specific pedestrians across multiple cameras,and is a subfield of image retrieval with broad application prospects in areas such as public safety and public security criminal investigation.However,there are often problems with similar pedestrian appearances,variable poses and obscured pedestrians in the images.The rapid development of deep learning and the emergence of convolutional neural networks have promoted the development of person reidentification.This paper explores the key issues in the field of person re-identification in the context of deep learning,and carry out in-depth research around the potential relationship between local and global features and multi-scale feature fusion,and do the following work:1.This paper proposes a person re-identification method based on the fusion of global and local correlation features.Firstly,in order to better distinguish similar local features of different samples,an improved relational network structure is proposed to obtain the relationship between neighbouring local features in the global picture by combining specific regional features with global features and constructing a combination of correlated features of different granularity to mine the potential semantic information they represent.Secondly,another contrast pooling branch obtains the maximum pooled features by contrasting pooled features with complementary information,making pedestrians more distinguishable through this correlation.The attention module based on spatial association is also introduced to make full use of the global contextual information of the network while mining the local association information of the features,and finally the global branch is added as an effective information supplement.In other words,the model learns more discriminative features through the collaboration of the three branching structures,while effectively combining the weighted fusion of the three loss functions.Rank-1 improved by 2.6% on CUHK03-Labeled and 3.5%on CUHK03-Detected dataset.2.This paper proposes a multi-scale learning person re-identification method based on the Transformer,which combines CNN and Transformer for application in the field of person re-identification.Firstly,in order to extend the remote dependencies,the JPM module is introduced into the Transformer and the improved Transformer is embedded into a convolutional neural network to better establish connections between long-range pixel points,thus compensating for the shortcomings of the convolutional neural network.In addition,the local branching incorporates an improved two-layer pyramid structure for extracting multiscale image local features,while combining the Non-local attention mechanism and orthogonal regularisation to effectively reduce the correlation of features between different channels and improve the performance of the network.Finally the image features extracted by the Transformer learning branch and the feature pyramid branch are supervised using hardsample sampling triad loss and cross-entropy loss.The m AP reached 89.2% and 80.8% on the Market1501 and Duke MTMC-re ID datasets,respectively,and 95.5% and 90.5% on Rank-1,respectively.