| 孙 伟,赵宇煌,张小瑞,刘轩诚.基于弱监督注意力和知识共享的车辆重识别[J].电子测量与仪器学报,2023,37(9):179-189 |
| 基于弱监督注意力和知识共享的车辆重识别 |
| Weakly supervised attention and knowledge sharing for vehicle re-identification |
| |
| DOI： |
| 中文关键词: 车辆重识别 弱监督 迁移学习 注意力机制 |
| 英文关键词:vehicle re-identification weak supervision transfer learning attention mechanism |
| 基金项目:国家自然科学基金(62376128,62272236)、江苏省自然科学基金(BK20191401, BK20201136)、江苏省研究生科研与实践创新计划项目(SJCX23_0380)、大学生创新创业训练项目(XJDC202110300601, 202010300290, 202010300211, 202010300116E)资助 |
|
|
| Author | Institution |
| Sun Wei | 1. College of Automation, Nanjing University of Information Science & Technology,2. Jiangsu CollaborativeInnovation Center on Atmospheric Environment and Equipment Technology, Nanjing University of Information Science &Technology |
| Zhao Yuhuang | 1. College of Automation, Nanjing University of Information Science & Technology |
| Zhang Xiaorui | 2. Jiangsu Collaborative Innovation Center on Atmospheric Environment and Equipment Technology, Nanjing University of Information Science &Technology,3. College of Computer and Information Engineering, Nanjing Tech University,4. Wuxi Research Institute, Nanjing University of Information Science & Technology,5. College of Computer and Software, Nanjing University of Information Science & Technology |
| Liu Xuancheng | 1. College of Automation, Nanjing University of Information Science & Technology |
|
| 摘要点击次数: 984 |
| 全文下载次数: 885 |
| 中文摘要: |
| 为解决弱监督车辆重识别方法中标签不准确和背景干扰下预定义局部区域难以获得的问题,提出一种基于弱监督注意
力和知识共享的车辆重识别网络。 在弱监督注意力模块(WAM)中使用弱监督的方法生成车辆部件掩模,通过部件通道对齐步
骤使得该模块在复杂背景下也能自适应地进行特征对齐。 针对弱监督方法中标签准确性不高导致 WAM 模块生成部件掩模不
稳定的问题,在局部分支中构建了知识共享模块。 该模块利用迁移学习从 WAM 模块中提取车辆部件特征,并进行多尺度部件
特征提取,防止了不稳定的车辆部件掩模生成。 通过实验,mAP、CMC@ 1 和 CMC@ 5 分别达到了 82. 12%、98. 50%和 99. 12%,
优于现有的方法,验明该方法的有效性。 |
| 英文摘要: |
| In order to solve the problem that the label is not accurate and the background interference makes it difficult to obtain the
predefined local area in the weak supervision vehicle re-identification method. A vehicle re-identification network based on weak
supervised attention and knowledge sharing is proposed. In the weak-supervised attention module (WAM), the weak-supervised method
is used to generate the vehicle component mask, and the component channel alignment step enables the module to perform adaptive
feature alignment under complex background. Aiming at the problem that the mask of WAM module is unstable due to the low accuracy
of labels in weak supervision method, a knowledge sharing module is constructed in local branches. The module uses migration learning
to extract vehicle component features from WAM module, and performs multi-scale component feature extraction to prevent unstable
vehicle component mask generation. Through experiments, mAP, CMC@ 1 and CMC@ 5 reached 82. 12%, 98. 50% and 99. 12%,
respectively, which are better than the existing methods and verify the effectiveness of this method. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
