基于残差注意和非对称损失的行人属性识别  

Recognition of Pedestrian Attributes Based on ResidualAttention and Asymmetric Loss

作　　者：胡红梅 张丽红[1] HU Hongmei;ZHANG Lihong(College of Physics and Electmnic Engineering,Shanxi University,Taiyuan 030006,China)

机构地区：[1]山西大学物理电子工程学院,山西太原030006

出　　处：《测试技术学报》2023年第2期99-105,共7页Journal of Test and Measurement Technology

基　　金：山西省研究生创新资助项目(2021Y154);山西省高等学校教学改革创新资助项目(J2021086)。

摘　　要：针对目前行人属性识别存在着复杂样本识别精度较低和属性数据集中属性分布不平衡的问题,本文提出一种基于残差注意的行人属性识别网络。该网络采用Resnet50作为骨干网络提取出具有语义信息的行人属性特征,并采用属性类别残差注意网络结构关注属性存在的关键区域且挖掘不同属性类别之间的内部联系。同时采用归一化和非对称的加权损失策略降低行人属性样本分布不平衡的影响,加快模型收敛速度并提高属性识别精度。在行人属性公共数据集PETA和PA100K上进行实验,实验结果表明,该方法在公共数据集PETA上获得的平均识别精度为87.32%,在公共数据集PA100K上可以获得79.75%的识别精度,与其他行人属性识别方法相比具有明显优势。In order to solve the problems of low recognition accuracy of complex samples and the unbalanced distribution of attributes in current pedestrian attribute recognition,this paper proposes a pedestrian attribute recognition network based on residual attention.The network uses Resnet50 as the backbone network to extract pedestrian attribute features with semantic information,and uses the attribute category residual attention module to focus on the key areas where the attributes exist and explore the internal connections between different attributes categories.At the same time,batch normalization(BN)and asymmetric weighted loss strategies are used to reduce the impact of the imbalance of pedestrian attribute samples,accelerate model convergence speed and improve the accuracy of attribute recognition.Experimental results on the public datasets of pedestrian attributes PETA and PA100K show that,the average recognition accuracy of this method can reach 87.32%on PETA and 79.75%on PA100K,which has obvious advantages compared with other pedestrian attribute recognition methods.

关 键 词：属性类别 残差注意 非对称损失 行人属性识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]