基于多分块三重损失计算的行人识别方法

Method based on the multi-block triplet loss calculation of person re-identification

提出一种基于多分块三重损失计算的行人识别方法,来解决行人识别准确率不高的问题。与其他方法相比,本文设计了一种多分块三重损失计算的方式来对提取的特征进行反馈计算。具体来说,首先,三个行人图像输入到分割层当中,各自分为上中下三块,并在对应的网络当中进行特征提取;其次,在经过两层卷积层提取行人特征之后,每一个行人的上中下三块都有对应的三重损失计算,并根据损失值进行反馈调节;最后,将所有的行人图像进行融合,并继续在网络当中训练,通过分类器进行分类输出,在测试当中根据输出值判别行人对是否属于同一个行人。实验表明,文章提出的方法在公开库数据库当中与其他传统方法相比显著提高了行人识别准确率。

In this paper,a new method based on multi-block triple-loss calculation of person re-identification is proposed to solve the problem of low pedestrian recognition accuracy.Multi-block triple loss calculation is utilized to carry out feedback calculation of the extracted features.Firstly,three pedestrian images are taken as the input of the slice layer and divided into upper,middle and lower blocks,respectively.The features are extracted in the corresponding networks.Secondly,the upper,middle and lower parts of each pedestrian use the corresponding triple loss calculation,and feedback is adjusted according to the loss value after extracting the pedestrian features from the two convolution layers.Finally,all the pedestrian images are fused and trained in the network.The output is classified and the pedestrian pairs are judged to the same pedestrian according to the output values.Experiments show that the method proposed provides higher recognition accuracy than other traditional methods in the open databases.