吉首大学学报(自然科学版) ›› 2021, Vol. 42 ›› Issue (6): 15-22.DOI: 10.13438/j.cnki.jdzk.2021.06.003

• 计算机 • 上一篇    下一篇

基于多尺度和小波空间注意力的车辆重识别

廖光锴,张正,牛一博,宋治国   

  1. (1.吉首大学信息科学与工程学院,湖南 吉首 416000;2.吉首大学物理与机电工程学院,湖南 吉首416000)
  • 出版日期:2021-11-25 发布日期:2022-01-21
  • 通讯作者: 宋治国(1984—),男,湖南保靖人,吉首大学物理与机电工程学院讲师,博士,主要从事目标检测、跟踪和识别研究.
  • 基金资助:
    国家自然科学基金资助项目(32060238)

Vehicle Re-Identification Based on Multi-Scale and Wavelet Spatial Attention

LIAO Guangkai, ZHANG Zheng, NIU Yibo, SONG Zhiguo   

  1. (1.School of Information Science and Engineering, Jishou University, Jishou 416000, Hunan China;2. School of Physics and Electromechanical Engineering, Jishou University, Jishou 416000, Hunan China)
  • Online:2021-11-25 Published:2022-01-21

摘要:针对基于多尺度的车辆重识别模型缺乏提取细节特征不足的问题,设计了一个融合多尺度的车辆特征和小波注意力机制的重识别模型.首先,将空间小波注意力模块镶入到模型中,能使网络获得更多细节特征;其次,提出一种阶梯融合网路,该网络对不同尺度层的特征进行分层融合,提高了模型聚合全局特征的能力;最后,使用TriHard和CrossEntropy Loss损失函数构造出车辆识别的目标函数.该模型在VeRi数据集和VehicleID数据集上与一些优秀模型的比较实验结果表明,将空间小波注意力嵌入到多尺度网络中能明显提高车辆重识别精确度.

关键词: 车辆重识别, 空间注意力机制, 小波变换, 多尺度, 特征融合

Abstract: For the problem that the multi-scale vehicle re-identification model is short of extracted detail features, a re-identification model combining multi-scale vehicle features and wavelet attention mechanism is designed.  Firstly, spatial wavelet attention module is embedded into the model to obtain more details for the network. Secondly, a hierarchical fusion network is proposed, which integrates the features of different scales to improve the ability of global feature aggregation. Finally, TriHard and CrossEntropy Loss functions are used to construct the object function of vehicle identification. Experimental results of this model compared with other excellent models on VeRi and VehicleID datasets show that vehicle re-identification accuracy can be improved by embedding spatial wavelet attention into multi-scale networks.

Key words: vehicle re-identification, spatial attention mechanism, wavelet transform, multi-scale, characteristics , fusion

公众号 电子书橱 超星期刊 手机浏览 在线QQ