基于双稀疏正则的图像集距离学习

发布时间：2019-07-21 14:36

【摘要】：随着视频采集和网络传输技术的快速发展以及个人移动终端设备的广泛使用,大量图像数据以集合形式存在.集合内在结构的复杂性使得如何度量集合间距离成为图像集分类的一个关键问题.为了解决这一问题,提出了一种基于双稀疏正则的图像集距离学习框架(double sparse regularizations for image set distance learning,简称DSRID).在该框架中,两集合间距离被建模成其对应的内部典型子结构间的距离,从而保证了度量的鲁棒性和判别性.根据不同的集合表示方法,给出了其在传统的欧式空间以及两个常见的流形空间,即对称正定矩阵流形(symmetric positive definite matrices manifold,简称SPD manifold)和格林斯曼流形(Grassmann manifold)上的实现.在一系列的基于集合的人脸识别、动作识别和物体分类任务中验证了该框架的有效性.
[Abstract]:With the rapid development of video capture and network transmission technology and the wide use of personal mobile terminal equipment, a large number of image data exist in the form of collections. Because of the complexity of the internal structure of sets, how to measure the distance between sets has become a key problem in image set classification. In order to solve this problem, a distance learning framework (double sparse regularizations for image set distance learning, for image sets based on double sparse regularity is proposed, which is called DSRID). For short. In this framework, the distance between the two sets is modeled as the distance between its corresponding internal typical substructures, thus ensuring the robustness and discrimination of the measurement. According to different set representation methods, its implementation on traditional Euclidean spaces and two common manifolds, namely symmetric positive definite matrix manifolds (symmetric positive definite matrices manifold, for short SPD manifold) and Greensmann manifolds (Grassmann manifold), is given. The effectiveness of the framework is verified in a series of set-based face recognition, action recognition and object classification tasks.
【作者单位】：交通数据分析与挖掘北京市重点实验室(北京交通大学);河北农业大学信息科学与技术学院;
【基金】：国家自然科学基金(61370129,61375062,61632004) CCF-腾讯犀牛鸟基金(RAGR20150116) 长江学者和创新团队发展计划(IRT201206) 保定市科学技术和知识产权局项目(16ZG026)~~
【分类号】：TP391.41

【相似文献】