基于暗光增强和细粒度特征提取的行人重识别  

Person re-identification based on dim light enhancement and fine-grained feature extraction

作　　者：杨靖 谷灵康 夏周祥 YANG Jing;GU Lingkang;XIA Zhouxiang(School of Computer and Information,Anhui Polytechnic University,Wuhu 241000,China)

机构地区：[1]安徽工程大学计算机与信息学院,安徽芜湖241000

出　　处：《陕西理工大学学报(自然科学版)》2025年第2期81-90,共10页Journal of Shaanxi University of Technology:Natural Science Edition

基　　金：安徽省重点实验室开放研究基金项目(KLMVI-2023-HIT-07);安徽省教育厅高校优秀青年人才支持计划重点项目(gxyqZD2019052)。

摘　　要：针对现实环境中由于光照变化和色差等因素导致的行人重识别准确率低的问题,设计了基于暗光增强和细粒度特征提取的行人重识别模型。首先利用暗光增强模块对输入的行人图像进行图像增强,优化图像信息,之后将图像输入改进后的Vision Transformer模型,该模型引入条件位置编码和拼图补丁促使网络关注行人图像的细粒度特征信息,增强对行人特征的提取能力。最后设计出DAD-T模块减少模型复杂度,缓解性能下降的问题。以此在性能无损的情况下提高训练效率,加快训练速度。在DukeMTMC-ReID、MSMT17和Market-1501三个行人重识别公开数据集上进行实验,并与主流网络模型进行比较。结果表明,所提模型的Rank-1与mAP指标相比当前主流模型有所提升,具有较高的识别准确率。A person re-identification model based on dim light enhancement and fine-grained feature extraction was designed to address the issue of low accuracy in person re-identification caused by factors such as lighting changes and color differences in the real environment.First,the dim light enhancement module is used to enhance the input person image and optimize the image information.Then the image is input into the improved transformer encoder,which introduces conditional positional encoding conditional positional encoding and jigsaw patches module.It prompts the network to pay attention to the fine-grained feature information of person images and enhances the ability to extract person features.Finally,the DAD-T module was designed to alleviate model performance degradation and reduce complexity.This improves training efficiency and speeds up training without loss of performance.Experiments were conducted on three public person re-identification data sets:MSMT17,DukeMTMC-ReID and Market-1501,and compared with mainstream network models.The results show that the Rank-1 and mAP indicators of the model proposed in this article have improved compared with the current mainstream models,and it has a high recognition accuracy.

关 键 词：暗光增强 细粒度 Vision Transformer模型 行人重识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]