The Research Of Person Re-identification Based On Multi-granularity Feature Fusion And Local Information Enhancement

Person re-identification is a technology that retrieves a given pedestrian image target across monitoring devices.It has been a key problem in the field of computer vision recently,and is widely used in intelligent security,human-computer interaction,e-commerce and other fields.Pedestrian images under surveillance have problems such as difference of perspective,change of posture,object occlusion,etc.This requires that person re-identification model can capture effective content for pedestrian identification.Effective pedestrian image characteristics can be extracted by using convolutional neural network.However,the global salient information of the image only pays attention to the overall situation of the image and is easily disturbed by irrelevant factors such as the external complex environment.How to make the network model pay more attention to the details that are more conducive to the identification through local information enhancement is the key problem to be solved in the person re-identification task.Based on this problem,this thesis mainly carries out the following improvements and innovations.(1)This thesis proposes a multi-granularity feature fusion network combining stepped feature extraction with branching attention to improve the performance of person re-identification task.In recent years,many researchers have used the method of horizontal uniform block segmentation to extract local features,but this method ignores some important information between blocks and fails to give full play to the advantages of local features.The segmenting of feature graph by ladder type can strengthen the connection between blocks and avoid the loss of edge information,so as to realize the local information enhancement to some extent.At the same time,branch attention is used to assign different weights to each branch,so that the network model can focus on the more important branches.In the design of network structure,a multi-branch network is used to combine the global and local features to achieve the coordination and unification of multi-granularity feature.Experimental results on public datasets show that the proposed method can obtain more discriminative features.(2)This thesis proposes a multi-scale adaptive local attention method that integrates multi-classes of attention to further obtain local salient information in person re-identification images.Attention mechanism is an effective method to enhance the attention degree of network model to the key information of image.In the past,the attention mechanism method in the person re-identification task aims to enhance the overall characteristics of the pedestrian.The proposed multi-scale adaptive attention module processes local features and adapts to images of different scales by adaptively adjusting the size of the receptive field.At the same time,channel attention and spatial attention are combined to screen out the important features of the image and use them for the final recognition.Finally,the proposed method is compared with several mainstream methods.Finally,after comparing with many mainstream methods,the network model designed in this thesis can effectively improve the accuracy of person re-identification task,which proves the superior performance of this method.To sum up,the method in this thesis mainly studies the local features of pedestrian images,seeks for appropriate methods to realize local image information enhancement,and takes it as an important supplement to global features in the form of multi-granularity.Finally,the excellent results of the proposed method were verified by experiments on three public pedestrian re-recognition datasets.