Direct extraction and fusion of image pixel features for face identity verification

In this paper, we propose to adopt an entry-wise Hadamard-Schur (HS) product to extract facial information directly from raw face image pixels. With the goal of extracting relevant pixel features from each image, four different HS projection matrices which consist of directional binary patterns are proposed. Subsequently, the match score outputs from these four HS projections are fused via a total error rate minimization for performance enhancement. Our empirical experiments show encouraging performance in terms of verification accuracy.