无线电工程2025，Vol.55Issue(3)：500-510,11.DOI:10.3969/j.issn.1003-3106.2025.03.006
基于跨阶段级联与多尺度注意力特征的行人重识别
Person Re-Identification Based on Cross-stage Cascade and Multi-scale Attention Feature
摘要
Abstract
As an instance-level recognition problem,the same pedestrian image in Person Re-Identification(ReID)may have large differences due to changes in viewing angle conditions,resulting in similarity discrimination heavily dependent on diverse features at different spatial scales.However,the ReID method based on Deep Convolutional Neural Network(DCNN)has the problem of easy loss of shallow basic information,and features of a single scale cannot fully describe the global relationship of pedestrians.Therefore,in order to obtain diverse pedestrian features,a ReID method based on Cross-Stage Cascade(CSC)and multi-scale attention features is proposed.A CSC structure is designed,combined with a Contextual Transformer Attention Block(CoT)mechanism,to mine and fuse shallow detail features of different stages,and use it as a global prior for deep abstract semantics.Then,the multi-scale convolution operation is used to process deep global semantic features,so that the model can learn pure global features and enhance the convolution model's ability to mine global feature relationships.Experiments on multiple data sets show that the network can effectively improve the accuracy of ReID.关键词
行人重识别/卷积神经网络/跨阶段级联/多尺度卷积/特征融合Key words
ReID/CNN/CSC/multi-scale convolution/feature fusion分类
计算机与自动化引用本文复制引用
沈宇慧,孟鑫,郭随平,郭业才..基于跨阶段级联与多尺度注意力特征的行人重识别[J].无线电工程,2025,55(3):500-510,11.基金项目
国家自然科学基金(61673222) （61673222）
江苏省研究生实践创新计划(SJCX22_0333) （SJCX22_0333）
南京信息工程大学研究生创新实践项目(WXCX202013)National Natural Science Foundation of China(61673222) （WXCX202013）
Jiangsu Province Graduate Practice and Innovation Program(SJCX22_0333) （SJCX22_0333）
Graduate Innovation Practice Project of Nanjing University of Information Science and Technology(WXCX202013) （WXCX202013）
