Recent research has demonstrated the effectiveness of linear representation (i.e., sparse representation, group sparse representation and collaborative representation) for face recognition and other vision problems. However, this linear representation assumption does not consider the non-linear relationship of samples and limits the usage of different features with non-linear metrics. In this paper, we present some insights of linear and non-linear representation-based classifiers. First, we present a general formulation known as kernel collaborative representation to encompass several effective representation-based classifiers within a unified framework. Based on this framework, different algorithms can be developed by choosing proper kernel functions, regularization terms, and additional constraints. Second, within the proposed framework we develop a simple yet effective algorithm with squared ℓ2-regularization and apply it to face recognition with local binary patterns as well as the Hamming kernel. We conduct numerous experiments on the extended Yale B, AR, Multi-PIE, PloyU NIR, PloyU HS, EURECOM Kinect and FERET face databases. Experimental results demonstrate that our algorithm achieves favorable performance in terms of accuracy and speed, especially for the face recognition problems with small training datasets and heavy occlusion. In addition, we attempt to combine different kernel functions by using different weights in an additive manner. The experimental results show that the proposed combination scheme provides some additional improvement in terms of accuracy.
Bibliographical noteFunding Information:
This work was supported by the China Postdoctoral Science Foundation under Grant 2014M551085 , the Fundamental Research Funds for the Central Universities under Grant DUT14YQ101 and DUT13RC(3)105, the Natural Science Foundation of China (NSFC) under Grants 61472060 , the joint Foundation of China Education Ministry , and by the China Mobile Communication Corporation under Grant MCM20122071 .
© 2015 Elsevier Ltd. All rights reserved.
All Science Journal Classification (ASJC) codes
- Signal Processing
- Computer Vision and Pattern Recognition
- Artificial Intelligence