融合GAT与Transformer的行人重识别方法  

Person re-identification method combining GAT and Transformer

作　　者：庄爽 宋建辉 刘鑫 ZHUANG Shuang;SONG Jianhui;LIU Xin(Shenyang Ligong University,Shenyang 110159,China)

机构地区：[1]沈阳理工大学自动化与电气工程学院,辽宁沈阳110159

出　　处：《通信与信息技术》2025年第2期17-22,共6页Communication & Information Technology

基　　金：辽宁省教育厅高等学校基本科研项目(项目编号:LJKZ0275);辽宁省属本科高校基本科研业务费专项资金资助;沈阳市中青年科技创新人才支持计划项目(项目编号:RC210247)。

摘　　要：针对行人重识别任务中,同一类行人在特征空间聚合度不佳,以及系统识别精度不高的问题,提出了一种结合图注意力神经网络(Graph Attention Network,GAT)与Transformer的行人重识别模型(Fusion of graph neural networks with the Transformer,FGT)。通过将Transformer的全局信息处理能力与改进GAT的局部特征提取能力相结合,为模型带来更深层次的信息理解和更广泛的特征表达能力。利用改进GAT分析输入图像之间的联系与结构特征,增强结构化数据的处理;通过改进深度可分离卷积并行去噪(Depthwise Separable Convolution Parallel Denoising,DCD)模块,减少噪声对特征提取的干扰;引入深度监督(Deep Supervision,DS)来避免梯度消失问题,促进网络快速收敛。在Market-1501、DukeMTMC-ReID、CUHK03和MSMT17数据集上的实验结果证明,本模型有效地增强了行人特征聚合度,提高了重识别准确率。In the person re-identification task,the convergence degree of the same type of person in the feature space is not good,and the recognition accuracy of the system is not high.In order to solve the problem,a person re-identification method(Fusion of graph neural networks with the Transformer,FGT)combining Graph Attention Network(GAT)and Transformer is proposed.By combining the global information processing ability of Transformer with the improved local feature extraction ability of GAT,it brings deeper information understanding and wider feature expression ability to the model.The improved GAT was used to analyze the relationship and structural features between input images to enhance the processing of structured data.The Depthwise Separable Convolution Parallel Denoising(DCD)module is improved to reduce the interference of noise on feature extraction.Deep Supervision(DS)is introduced to avoid the gra-dient disappearance problem and promote the rapid convergence of the network.The experimental results on the Market-1501,DukeMT-MC-ReID,CUHK03,and MSMT17 datasets demonstrate that this model effectively enhances the degree of person feature aggregation and improves the accuracy of re-identification.

关 键 词：行人重识别 图注意力神经网络 TRANSFORMER 深度可分离卷积并行去噪 深度监督 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]