摘要
提出了一种基于深度学习自适应权重分配监控视频行人再识别方法。基于验证损失所反映的行人属性训练难度,结合行人属性与行人类别对应关系的信息熵,求解属性对分类的贡献率,自适应求解行人属性多任务分类的训练损失权重,解决多任务分类时分配相同的损失权重造成的负转移问题,以提高每个任务学习器的泛化能力以及对行人类别判别的泛化能力。利用已有数据集中行人属性与行人类别的映射关系,根据已训练好的模型求解属性概率,结合条件概率判别行人类别,克服全局行人再识别中无法识别网络视角变化造成的行人外观剧烈变化的问题。与同类方法在不同公开数据测试集上的客观定量进行对比,结果表明所提方法有效、可行。
This paper proposes a method to monitor video-based pedestrian re-identification based on adaptive weight assignment using deep learning. The contribution rate of the pedestrian attribute to the classification is calculated based on the training difficulty of the pedestrian attribute reflected by verification loss along with the correspondence information entropy of the pedestrian attribute and pedestrian category .The training loss weight of the pedestrian attribute multi-task classification is adaptively solved. The negative transfer problem caused by the same loss weight is assigned to improve the generalization abilities of each task learner and pedestrian re-identification. The trained model solves the attribute probability and combines the conditional probability to discriminate the pedestrian category using the mapping relationship between the pedestrian attribute and the pedestrian category in the existing data set, which overcomes the problem that cannot identify pedestrian category because of the dramatic change of the pedestrian appearance. Based on objective and quantitative comparison with similar methods on different public data test sets, the results show that the method is effective and feasible.
作者
李净
管业鹏
Jing Li;Guan Yepeng(School of Communication and Information Engineering, Sharighai University, Shanghai 200444, China;Key Laboratory of Advanced Display and System Application , University of Education,Shanghai University, Shanghai 200072, China)
出处
《激光与光电子学进展》
CSCD
北大核心
2019年第14期63-72,共10页
Laser & Optoelectronics Progress
关键词
图像处理
行人再识别
深度学习
自适应权重
行人属性
image processing
pedestrian re-identification
deep learning
adaptive weight
pedestrian attribute
