适应跨域的行人重识别算法  

Person re-identification algorithm adapted to cross-domain

作　　者：石祥滨 吴天宇 SHI Xiang-bin;WU Tian-yu(College of Computer Science,Shenyang Aerospace University,Shenyang 110136,China)

机构地区：[1]沈阳航空航天大学计算机学院,沈阳110136

出　　处：《沈阳航空航天大学学报》2022年第6期46-56,共11页Journal of Shenyang Aerospace University

基　　金：国家自然科学基金(项目编号:61170185)。

摘　　要：针对在跨域场景下,行人图片受到光照、色度变化、风格转换等因素影响,行人重识别学习得到的模型无法较好表达行人外观特征的问题,对常用的分类网络骨架ResNet作出修改,并结合IBN模块解决图像风格转换问题。同时提出随机灰度擦除的数据增强方案,在不改变原始图像结构信息的前提下,使用灰色补丁图像代替原始图像,有效降低光照对行人重识别模型的影响。其次,在模型训练中使用Circle Loss代替了SoftMax Loss和Triplet Loss的线性结合,并对Center Loss进行加权以加强特征的内聚性,同时结合Warm-Up和余弦退火策略动态改变学习率。最后,在不降低精度的前提下通过知识蒸馏方法学习得到较小的模型完成加速推理。实验表明,所提出的方法可以有效提升模型收敛速度,并在现实应用中完成快速跨域行人检索。In view of the cross-domain,pedestrian pictures are affected by factors such as illumination,chromaticity changes and style conversion,etc.,and the model obtained by person re-identification learning cannot well express the appearance characteristics of pedestrians.The commonly used classification network skeleton ResNet was modified,and IBN block was introduced to deal with the problem of image style conversion.At the same time,a data enhancement scheme of random grayscale erasure was proposed.On the premise of not changing the original image structure information,the gray patch image was used to replace the original image,which effectively improved the impact of illumination on the person re-identification model.Secondly,Circle Loss was used in model training to replace the linear combination of Soft Max Loss and Triplet Loss,and the Center Loss was added to enhance the cohesion of features,while combining warm-up and cosine annealing strategies to dynamically change the learning rate.Finally,the knowledge distillation method was used to learn a smaller model to complete accelerated inference without reducing the accuracy.Experiments show that this method can effectively improve the convergence speed of the model and complete fast cross-domain pedestrian retrieval in real applications.

关 键 词：行人重识别 IBN模块 随机灰度擦除 训练策略 跨域 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]