In this paper, we present a rich image representation which is robust to illumination, facial expression and scale variations. For this aim, firstly, we propose a novel dense local image representation method based on Walsh Hadamard Transform (WHT) called Local WHT (LWHT). LWHT is the application of WHT to each pixel of an image to decompose it into multiple components, called LWHT maps. Secondly, although LWHT maps are real valued images we propose a method to produce complex valued images from LWHT maps by pairing these maps. We utilize these complex valued image components to obtain Phase Magnitude Histograms (PMHs) in feature extraction stage. Experiments on FERET dataset show that LWHT outperforms Local Binary Patterns (LBP) and Local Gabor Binary Patterns. To further improve the recognition performance, we enhanced our basic method by dividing images into subregions and weighting them, applying cascaded LWHT, and reducing dimension of feature vectors by Block-based Whitened Principal Component Analysis (BWPCA). Experimental results show that the proposed algorithm considerably improves the Walsh-based face recognition and generate comparable results for LBP and Gabor based approaches.