基于颜色随机化和全相关注意力的跨模态行人重识别  

Cross-modal pedestrian re-recognition based on color randomization and full related attention

作　　者：余文涛 赵倩 季堂煜 Yu Wentao;Zhao Qian;Ji Tangyu(College of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 201306,China)

机构地区：[1]上海电力大学电子与信息工程学院,上海201306

出　　处：《国外电子测量技术》2023年第6期10-16,共7页Foreign Electronic Measurement Technology

基　　金：国家自然科学基金(61802250)项目资助。

摘　　要：针对跨模态行人重识别过程中,模态差异导致难以提取充分的辨别性身份特征的问题,提出一种颜色随机化数据增强算法,并设计了基于全相关注意力的双流多分支网络模型。模型以ResNet-50为骨干网络,首先,对输入样本进行颜色随机化处理,提高模型的颜色风格鲁棒性;采用双流网络,在网络浅层设置权重参数非共享模式,分别用于处理可见光图像和红外图像;其次,提出全相关注意力,从空间和通道维度获得不同像素的关联程度,提高模型对于结构信息的提取能力;最后,采用多分支结构提取多尺度全局特征和局部特征增强提取特征的判别性。实验结果表明,所提方法在SYSU-MM01数据集的全搜素测试模式下,Rank-1和平均精度均值(mAP)分别达到70.01%和67.40%,优于其他方法。To address the problem that modal differences between different modal images in the process of cross-modal pedestrian re-recognition make it difficult to extract sufficient discriminative features,this paper proposes a color randomization data pre-processing method and designs a dual-stream multi-branch network model based on full relevant attention.The model uses ResNet-50 as the backbone network.Firstly,the input image is pre-processed with color randomization to improve the color style robustness of the model;a dual-stream network with non-shared modes of weight parameters is set in the shallow layer of the network for processing visible and infrared images respectively;secondly,full related attention is proposed to obtain the degree of association of different pixels from the spatial and channel dimensions to improve the model's ability.Finally,the multi-branch structure is used to extract multi-scale global features and local features to enhance the discriminative power of the extracted features.The experimental results show that the proposed method achieves 70.01%and 67.40%for Rank-1 and mAP,respectively,under the full search element test mode of SYSU-MM01 dataset,which is better than other methods.

关 键 词：跨模态 行人重识别 注意力机制 图像增强 多分支结构 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]