International Journal of Applied Mathematics and Machine LearningAuthor's: Yinghao Deng, Qianjin Zhao, Shuzhi Su, Ruonan Zhang and Penglian Gao
Pages: [17] - [28]
Received Date: December 29, 2019
Submitted by: Jianqiang Gao.
DOI: http://dx.doi.org/10.18642/ijamml_7100122110
Canonical Correlation Analysis (CCA) is a classical feature learning method, which is widely used in image recognition, information fusion, and affective computing and so on. However, it is difficult for CCA to find nonlinear local sub-manifold structure hidden in the raw sample space. In view of this issue, locality preserving canonical correlation analysis (LPCCA) is proposed, which overcomes the preservation of local geometrical structure in CCA. However, LPCCA still does not consider the global Euclidean structure in the raw sample space. To solve this problem, we propose an elastic canonical correlation analysis method that preserves both local geometry structure and global Euclidean structure hidden in the raw sample space. The method is successfully applied to image recognition, and a large number of experimental results have showed the superiority of the method.
feature extraction, subspace learning, image recognition.
