This paper presents a novel method for face recognition. First, we generate the new image representation from the decorrelated hybrid color configurations rather than RGB color space via a learning algorithm. The learning algorithm, Principal Component Analysis (PCA) plus Fisher Linear Discriminant analysis (FLD), is able to derive the desired color transformation to generate a discriminating image representation that is optimal for face recognition. Second, we partition face image into some small patches, each of which can obtain its own color transformation, to reduce the effect of illumination variations. Thus, a patch-based novel image representation method is proposed for face recognition. Experiments on the Face Recognition Grand Challenge (FRGC) version 2 Experiment 4 show that the proposed method outperforms gray-scale image and some recent methods in face recognition.