摘要
针对因拍摄的行人图像模糊、遮挡、姿势、视角、颜色、风格和亮度不同等不良因素的影响,行人重识别任务难以提取具有判别力的特征,设计了一种基于IBN-Net和双池通道注意力模块的新颖网络IBNC-Net。以IBN-Net50-a作为骨干网络学习不因图像风格、颜色和亮度等外观变化而变化的特征;在不同的网络层嵌入双池通道注意力模块DPCAM,抑制无关特征,增强具有判别力的特征;引入广义平均池化GeM,通过模型训练自动调整池化尺度。为了验证提出的IBNC-Net方法的有效性,在三个流行的数据集上进行实验,包括Market1501、DukeMTMC-ReID和CUHK03。IBNC-Net模型的Rank-1分别达到了95.6%、91.2%和80.5%,mAP分别达到了89.1%、80.3%和79.4%,实验结果表明,所提方法能够有效提高行人重识别模型的性能。
Due to the influence of bad factors such as blurred pedestrian images,occlusion,posture,different angles of view,color,style and brightness,it is difficult to extract distinguishing features in person re-identification task.This paper designs a novel network IBNC-Net based on IBN-Net and dual pool channel attention module.Firstly,IBN-Net50-a is used as the backbone network to learn the features that do not change with the appearance changes such as image style,color and brightness.Secondly,the dual pool channel attention module(DPCAM)is embedded in different network layers to suppress irrelevant features and enhance distinguishing features.Then,the generalized mean pooling(GeM)is intro-duced,and the pool scale is automatically adjusted by the model through training.Finally,in order to verify the effective-ness of the proposed IBNC-Net method,experiments are conducted on three popular datasets,including Market1501,DukeMTMC-ReID and CUHK03.The proposed IBNC-Net model has reached 95.6%/89.1%,91.2%/80.3%and 80.5%/79.4%(mAP/Rank-1)respectively.Experimental results show that the proposed method can effectively improve the performance of person re-identification model.
作者
杨永胜
邓淼磊
张德贤
YANG Yongsheng;DENG Miaolei;ZHANG Dexian(College of Information Science and Engineering,Henan University of Technology,Zhengzhou 450001,China;Henan International Joint Laboratory of Grain Information Processing,Zhengzhou 450001,China)
出处
《计算机工程与应用》
CSCD
北大核心
2023年第17期143-151,共9页
Computer Engineering and Applications
基金
国家重点研发计划
河南省重大公益专项(201300311200)。
关键词
行人重识别
深度学习
注意力机制
计算机视觉
person re-identification
deep learning
attention mechanism
computer vision
