Novel methods for patch-based face recognition

Patch-based face recognition is a robust method which aims to tackle illumination changes, pose changes and partial occlusion at the same time. In this paper, we analyzed the effects of different dimension reduction and normalization techniques for patch-based face recognition. Apart from previously used dimension reduction methods such as DCT and PCA, we have applied NPCA and NNDA at the feature extraction stage. Following the feature extraction, feature fusion or decision fusion can be applied at the recognition stage. In this study, we have shown that decision fusion outperforms feature fusion which is previously used in patch-based face recognition. For decision fusion, we proposed novel method for calculating weights for the weighted sum rule. We obtained the highest recognition rates by applying NNDA for dimension reduction, norm division for normalization and VAWfor assigning weights to patches.