查宇飞,,
李运强,
张胜杰,
张园强
空军工程大学航空航天工程学院 ??西安 ??710038
基金项目:国家自然科学基金(61472442, 61773397, 61701524)，陕西省科技新星资助(2015kjxx-46)
详细信息
作者简介:陈兵：男，1994年生，博士生，研究方向为计算机视觉、行人重识别
查宇飞：男，1979年生，副教授，研究方向为计算机视觉及模式识别、目标检测、目标跟踪
李运强：男，1992年生，博士生，研究方向为计算机视觉、二值图像检索及人脸识别
张胜杰：男，1994年生，硕士生，研究方向为计算机视觉、二值图像检索
张园强：男，1994年生，硕士生，研究方向为计算机视觉、视觉目标跟踪
通讯作者:查宇飞 735754591@qq.com
中图分类号:TP391.41计量
文章访问数:993
HTML全文浏览量:317
PDF下载量:34
被引次数:0
出版历程
收稿日期:2018-02-09
修回日期:2018-07-17
网络出版日期:2018-07-23
刊出日期:2018-10-01
Shift-variant Similarity Learning for Person Re-identification
Bing CHEN,Yufei ZHA,,
Yunqiang LI,
Shengjie ZHANG,
Yuanqiang ZHANG
Aeronautics Engineering College, Air Force Engineering University, Xi’an 710038, China
Funds:The National Natural Science Foundation of China (61472442, 61773397, 61701524), Shanxi Science and Technology New Star Fund (2015kjxx-46)
摘要
摘要:行人重识别的精确度主要取决于相似性度量方法和特征学习模型。现有的度量方法存在平移不变性的特点，会增加网络参数训练的难度。现有的几种特征学习模型只强调样本之间的绝对距离而忽略了正样本对和负样本对之间的相对距离，造成网络学习到的特征判别性不强。针对现有度量方法的缺点该文提出一种平移变化的距离度量方法，能够简化网络的优化并能高效度量图像之间的相似性。针对特征学习模型的不足，提出一种增大间隔的逻辑回归模型，模型通过增大正负样本对之间的相对距离，使得网络得到的特征判别性更强。实验中，在Market1501和CUHK03数据库上对所提度量方式和特征学习模型的有效性进行验证，实验结果表明，所提度量方式性能更好，其平均精确率超出马氏距离度量6.59%，且所提特征学习模型也取得了很好的性能，算法的平均精确率较现有的先进算法有显著提高。
关键词:行人重识别/
平移变化/
相似性学习/
逻辑回归
Abstract:The accuracy of pedestrian re-recognition mainly depends on the similarity measure and the feature learning model. The existing measurement methods have the characteristics of translation invariance, which make the training of network parameters difficult. Several existing feature learning models only emphasize the absolute distance between sample pairs, but ignore the relative distance between positive sample pairs and negative sample pairs, resulting in a weak discriminant feature in network learning. In view of the shortcomings of existing measurement methods, a distance measurement method of translation change is presented, which can effectively measure the similarity between images. To overcome the shortcomings of the feature learning model, based on the proposed translation distance metric, a new logistic regression model with enlarged intervals is proposed. By increasing the relative distance between the positive and negative sample pairs, the network can get more discriminant features. In the experiment, the validity of the proposed measurement and the feature learning model is verified on the Market1501, CUHK03 database. Experimental results show that the proposed metric performs better than the Mahalanobis distance metric 6.59%, and the proposed feature learning algorithm also achieves good performance. The average precision of the algorithm is improved significantly compared with the existing advanced algorithms.
Key words:Person Re-identification/
Shift-variant/
Similarity learning/
Logistic regression
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=79714143-9552-4e06-9764-864ac376a482
