基于多粒度特征融合网络的行人重识别

Person re-identification based on multi-granularity feature fusion network

摘要:

针对复杂环境下行人细节特征不明显、姿态多变等 情况造成的行人重识别(person-reidentification,ReID)算法精度不高的问题,提出了一种基于多粒度特征提取与特征融合的ReID网络。首先,在主干网络输入和输出端采用两种粒度的划分方式获取图像的局部 特征。其次, 引入空间变换网络(spatial transformation network,STN)对输入全局图像进行空间对齐,对局部图像进行特征增强。最后 ,采用局部特征融合的方式来挖掘 特征之间的关联信息,提升模型对相似样本的识别能力。实验结果表明,所提方法在多个数 据集上均取得了良好的识别效果。 在Market-1501数据集上的平均查准率(mean average precision,mAP)和首次查准率(Rank-1)分别为84 .87%和94.45%,通过和目前主流的ReID算法相比,本文所提方法具有更优的识别效果。

Abstract:

In order to solve the problem of low a ccuracy of person re-identification (ReID) algorithm in complex environment caused by unc lear detail features and changeable attitudes of pedestrians,a ReID network based on multi-granularity feature extraction and feature fusion is proposed.Firstly,tw o granularity partitioning methods are used to obtain the local features of the image at the input and output ends of the b ackbone network.Secondly,spatial transformation network (STN) is introduced to align the global image and enhance the local image.Finally,local feature fusion is used to mine the correlation information between features to improve t he model′s ability to recognize similar samples.Experimental results show that the proposed method achieves good recogn ition performance on multiple datasets. The mean average precision (mAP) and first accuracy (Rank-1) of the market-1501 data set are 84.87% and 94.45%,respectively. Compared with the current mainstream ReID algorithms,the prop osed method has better recognition effect.