Robust kernel-based learning for image-related problems

Robustness is one of the most critical issues in the appearance-based learning techniques. This study develops a novel robust kernel for kernel machines, and consequently improves their robustness in resisting noise for solving the image-related learning problems. By incorporating a robust ρ-function to reduce the influence of outlier components, this kernel gives more reasonable kernel values when images are seriously corrupted. The authors incorporate the proposed kernel into different kernel-based approaches, such as support vector machine (SVM) and kernel Fisher discriminant (KFD) analysis, to validate its performance on various visual learning problems of face recognition and data visualisation. Experimental results indicate that the proposed kernel can provide the superior robustness to the classical approaches.