一种利用目标结构关系增强的行人重识别方法

A structural relationship enhanced person re-identification method

行人重识别旨在寻找与查找图像匹配的最佳目标图像,在视频目标搜寻中具有极大的应用意义。由于现有行人重识别方法忽略了对行人目标特征结构关系的挖掘,本文提出一种利用行人目标结构关系的结构增强可堆叠注意力模块。该模块通过计算某特征通道与其余通道之间的目标特征结构关系响应值来增强特征图对结构信息的表示能力。通过在经典行人重识别数据集Market-1501及CUHK03数据集上进行实验来验证模块有效性。实验结果表明,在ResNet网络中嵌入该可堆叠模块可学习到比先前方法更具区分性的特征表达,进而提高了行人重识别的准确率。

Person re-identification aims to find the image that matches the query image most, which is significant in target retrieval from video.Previous approches lack of mining for structural relationship of target feature, structure-enhanced stackable attention module using structure-dependent feature is proposed. This module enhances the structural infomation in feature map by computing the structural relationship value between centain channel and the rest ones.Through plenty of experiment in Market1501 and CUHK03 datasets, which are classic benchmark in person Re-identification, by statcking the proposed module in ResNet, the SESNet notices the distinguishable region and achives competitive perfomance toward the privious approches.