摘要
针对现有的基于卷积神经网络的行人重识别方法对于遮挡和复杂背景引起的判别信息缺失问题,提出了一种基于多尺度卷积特征融合的行人重识别算法。在训练阶段,使用金字塔池化方法对卷积特征图进行分块和池化,获得包含全局特征和多尺度局部特征的多个特征向量;对每一个特征向量进行独立分类,并在各个分类的最后内积层上归一化权重和特征,以提升分类性能;最后使用梯度下降法优化全部的分类损失。在识别阶段,将池化后的多个特征向量融合成一个新向量,使用新向量在库中进行相似性匹配。在Market-1501、DukeMTMC-reID数据库上对所提算法的有效性进行实验验证。结果表明,本文模型提取的特征具有更好的识别效果,Rank-1精度和平均准确率也优于大多数先进算法。
Existing methods of person reidentification based on convolutional neural network lack discriminative information,due to occlusion and complex backgrounds.To solve these problems,a method based on multi-scale convolutional feature fusion is proposed herein.In the training phase,pyramid pooling is used to extract multiple eigenvectors containing global features and multi-scale local features for blocking and pooling of the convolutional feature map.Afterward,each feature vector is classified independently,and the weights and features on the last inner layer of each class are normalized to improve the classification performance.Finally,agradient descent algorithm is applied to optimize the sum of losses for each classification.In the recognition phase,pooled multiple feature vectors are concatenated into a new vector for similarity matching.The efficiency of the proposed algorithm is verified on datasets Market-1501 and DukeMTMC-reID,in which the results indicate that features obtained by the proposed model are more discriminative and that the Rank-1 accuracy and average accuracy are both better than most state-of-the-art algorithms.
作者
徐龙壮
彭力
Xu Longzhuang;Peng Li(Engineering Research Center of Internet of Things Technology Applications of the Miyiistry of Education,School of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China)
出处
《激光与光电子学进展》
CSCD
北大核心
2019年第14期213-219,共7页
Laser & Optoelectronics Progress
基金
国家自然科学基金(61873112)
教育部-中国移动科研基金(MCM20170204)
江苏省博士后科研资助计划(1601085C)
关键词
机器视觉
卷积神经网络
金字塔池化
多尺度特征融合
行人重识别
权重与特征归一化
machine vision
convolutional neural network
pyramid pooling
multi-scale feature fusion
person reidentification
weight and feature normalization
