基于属性敏感卷积神经网络的行人属性分类  

Pedestrian attribute classification based on label sensitive convolutional neural network

作　　者：王进[1] 黄超[1] 王科[1] 范磊 陈乔松[1] WANG Jin;HUANG Chao;WANG Ke;FAN Lei;CHEN Qiaosong(Chongqing Key Laboratory of Computational Intelligence,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学计算智能重庆市重点实验室

出　　处：《江苏大学学报（自然科学版）》2019年第4期431-438,共8页Journal of Jiangsu University：Natural Science Edition

基　　金：国家自然科学基金资助项目(61203308,61309014,61403054);重庆市基础与前沿研究计划项目(cstc2014jcyjA40001);重庆教委科学技术研究项目(KJ1400436)

摘　　要：针对行人属性分类受行人属性不均衡影响的问题,提出了一种基于属性敏感卷积神经网络的行人属性分类方法.首先调整现有的卷积神经网络结构,通过融合正反通道激活模块的使用,使模型能够感知更加详细的行人属性;其次引入属性不均衡损失函数,根据属性的不均衡比例自适应更新网络权重,利用误差的反向传播对少类属性增加其权值,提升模型对少类属性的敏感;最后在PETA数据集上,对54个属性进行了分类试验.结果表明:相比MLCNN等方法,新方法在36分类任务上取得了提升;在平均准确度、平均召回率和平均AUC上,分别提升2.13%,2.38%和1.19%.To solve the problem that the classification of pedestrian attributes was affected by the imbalanced pedestrian attributes,a pedestrian attribute classification was proposed based on label sensitive convolutional neural network(CNN).The existing convolutional neural network structure was adjusted.By integrating the positive and negative channel activation modules,the model could perceive more detailed pedestrian attributes.The attribute imbalance loss function was introduced,and the network weights were adaptively updated according to the imbalance ratio of attributes.The back propagation of error was used to increase the weight of small-class attributes,and the sensitivity of model to the small-class attributes was improved.54 attributes were classified on the PETA data set.The results show that compared with MLCNN and other methods,the new method is improved for 36 classification tasks.The average accuracy,the average recall and the average AUC are increased by 2.13%,2.38%and 1.19%,respectively.

关 键 词：卷积神经网络 行人属性分类 属性敏感 属性不均衡 误差反向传播 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]