Research On Person Re-Identification Algorithm Based On Attention Mechanism And Unsupervised Learning

The purpose of person re-identification is to retrieve a specific person based on visual appearance in the cross-camera monitoring network.It is one of the core technologies of intelligent monitoring and it has various important application scenarios,such as video surveillance,human behavior analysis,and multitarget tracking.In recent years,more and more researchers apply convolutional neural network to person re-identification task to extract pedestrian features.However,the existing large appearance variations caused by background noises,different viewpoints,and illumination conditions.These challenges increase the difficulty of person re-identification task.Moreover,due to the difficulty of making person re-identification data and the weak generalization ability of a large number of current person re-identification models,these problems seriously hinder the wide application of person re-identification in real life.To solve the above problems,two different algorithms to solve the corresponding problems in this paper,respectively.The main research contents are as follows:(1)In order to extract person features with more representation ability,and solve the problem of large changes in pedestrian appearance caused by background noise,camera-style changes and illumination changes.This paper proposes a Multi-information Fusion reinforced Global Attention(MIFGA)module,which can obtain the relation information between spatial positions and channels in the feature maps from many aspects,so as to better guide the learning of attention and promote the model to extract more discriminative features.Specifically,the MIFGA includes spatial attention MIFGA-S and channel attention MIFGA-C.In the spatial attention,there are more topology information because of more semantic information in the deep features,such as the topology information between pedestrian limb features.To mine the potential topology information in feature maps,we go further and propose the Self-Learning Graph Convolution Network(SLGCN).MIFGA-S fuses local-feature semantic information and spatial topological information to guide spatial attention learning.In the channel attention,we can focus on the more important channels by using the weighted superposition of channel pairs affinity to enhance the representation ability of feature maps.MIFGA-C fuses channel semantic information and channel affinity information to guide the learning of channel attention.A large number of comparative experiments and ablation experiments on public person re-identification data sets prove that the MIFGA proposed in this paper can effectively improve the performance of pedestrian re recognition model.(2)Due to the difficulty of making person re-identification data set and the weak generalization ability of the existing person re-identification model,person re-identification is difficult to be widely used in real life.Therefore,this paper adopts the setting of unsupervised person re-identification,which is no need to use the person re-identification data set with pedestrian ID mark.It greatly reduces the production cost of the data set.This paper proposes an Asymmetric Double Networks Mutual Teaching(ADNMT)framework which uses two asymmetric networks to generate pseudo-labels for each other by clustering.The pseudo-labels are then updated and optimized by alternate training.Specifically,ADNMT contains two asymmetric networks.One is a multiple granularity network which extracts pedestrian features of multiple granularities and corresponds to multiple classifiers,and the other is a conventional single branch backbone network which extracts pedestrian features and corresponds to a classifier.Furthermore,this paper designs the Similarity Compensation of Inter-Camera and Similarity Suppression of Intra-Camera according to the camera ID of pedestrian images to optimize the similarity measures.So as to reduce the impact of the change of camera style on the performance of unsupervised person re-identification.A large number of experiments show that ADNMT can solve the problem of unsupervised person re-identification better than other advanced unsupervised person re-identification algorithms.