Fusing Local Patterns of Gabor and Non-subsampled Contourlet Transform for Face Recognition

Gabor features have been demonstrated to be very effective for face representation. Recently, non-sub sampled contour let transform (NSCT), which is a newly developed multi-resolution analysis tool based on contour let transform, is also used in facial image processing. In fact, the two image decomposition methods are performed from two different angles. To exploit complementarity of these features, in this paper, we propose a new face representation based on fusing local patterns of Gabor and NSCT. Firstly, we decompose face images using Gabor and NSCT respectively. Then all decomposition images are encoded by local texture descriptors to combine. In order to extract efficient features for feature fusion, we propose local Gabor difference features (LGDF) and local contour let difference features (LCDF) to represent the texture of decomposition images. Thirdly, after fusing LGDF and LCDF, block-based Fisher's linear discriminant (BFLD) is utilized to further reduce the dimensionality and improve discriminative power of the proposed method. Experiments on public databases demonstrate that the proposed LGDF and LCDF are very effective and our approach outperforms many state-of-the-art methods.