行人属性识别:基于元学习的概率集成方法  

Pedestrian Attribute Recognition:Probabilistic Ensemble Learning Method Based on Meta-learning

作　　者：王文龙 张磊[1] 张誉馨 吴晓富[1] 张索非[2] WANG Wen-long;ZHANG Lei;ZHANG Yu-xin;WU Xiao-fu;ZHANG Suo-fei(School of Communication and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;School of Internet of Things,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003 [2]南京邮电大学物联网学院,江苏南京210003

出　　处：《计算机技术与发展》2022年第3期71-75,83,共6页Computer Technology and Development

基　　金：国家自然科学基金(61701252)。

摘　　要：行人属性识别(pedestrian attribute recognition,PAR)的目的是从输入图像中挖掘行人的属性信息。近年来,卷积神经网络(convolution neural network,CNN)的兴起在行人属性识别中获得了广泛的应用。现有的方法多采用属性不可知的视觉注意或启发式的身体部位定位机制来增强局部特征表达,而忽略了多模型集成所能够带来的提升,因此该领域内鲜少有集成算法的提出。为了进一步提高行人属性识别的性能,该文从CNN模型的预测概率角度入手,基于元学习提出了一种行人属性识别的概率集成算法(probabilistic ensemble learning method,PEM)。在行人属性识别数据集RAP上的实验结果表明,该算法随着模型有效数的增加表现出递增的平均准确度(mean accuracy)和F1值(F1-score),且评测结果均优于多个典型的行人属性识别算法。Pedestrian attribute recognition(PAR)aims to mine pedestrian attribute information from input images.In recent years,convolution neural network(CNN)has been widely used in PAR.Existing methods mostly use visual attention with unknown attributes or heuristic body part localization mechanism to enhance local feature expression,but ignore the improvement brought by multi-model integration.Therefore,there are few integration algorithms proposed in this field.In order to further improve the performance of PAR,we propose a probabilistic ensemble learning method based on meta-learning from the perspective of CNN model’s prediction probability.The experimental results on RAP dataset of pedestrian attribute recognition show that the proposed algorithm achieves improved mean accuracy(mA)and F1 score with the increase of models,which are better than several existing pedestrian attribute recognition algorithms.

关 键 词：卷积神经网络 行人属性识别 元学习 概率集成 行人检索 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]