Supplementary Materials: Kernel Sparse Representation with Pixel-level and Region-level Local Feature Kernels For Face Recognition

We compared the proposed KCDSRC algorithm with the KMTJSRC algorithm [3] on the Extended YaleB and the CMU-PIE databases. The proposed LBPh-KH kernel is used due to its best overall performance than the other kernels. The experiment is conducted under three conditions, illumination, random noise, and synthesized continuous occlusion, where the settings are the same as we have applied before. The KMTJ source code which is used in our experiment is available on the author’s homepage. The parameters of KMTJ used in the experiment are the same as the KCD. The recognition results along with their standard deviations are shown in Fig.s1. In each figure there are four curves, which stand for the KCDSRC method with LBPh-KH kernel on the Extended YaleB and the CMU-PIE (YaleKCD & PieKCD), and the KMTJSRC method with LBPh-KH method on the Extended YaleB and the CMU-PIE (YaleKMTJ & PieKMTJ) respectively. On the left of Fig.s1, it is the recognition results of the KCDSRC and KMTJSRC algorithms tested on original images of the YaleB and the CMU-PIE databases. We can see that the performance on the CME-PIE is better than on the YaleB, which is also occurred in the illumination part subsection 4.2. And the reason lies in the stronger illumination changes on the YaleB database compared to the CMU-PIE (see Fig.6). It can also be observed that the KCDSRC with LBPh-KH kernel performs better than the KMTJSRC with LBPh-KH on both the YaleB and the CMU-PIE databases. With