Image-Quality-Based Fusion Approach for Face Recognition

Discrimination, robustness and inexpensiveness in both terms of time and storage are the three most important properties of a good face recognition system. A recent feature descriptor called Patterns of Oriented Edge Magnitude (POEM) balances three concerns. However, this feature descriptor does not take account different lighting conditions on different regions of the given face and simply concatenates different regions on the given face to get the histogram sequence to represent the face, which will reduce the recognition accuracy. Motivated by these analyses, this paper presents a face recognition system by combining the robust illumination normalization, the efficient POEM feature descriptor and the multiple region feature fusion approach. This paper makes two main contributions: 1) it presents a simple and efficient preprocessing method to reduce the effect of varying illumination; 2) it proposes a novel image-quality-based fusion approach by incorporating the histogram sequence estimated from different regions on the given face. The experiment results on the Extend Yale Face Database B show that the proposed face recognition system using image-quality-based fusion approach has better performance than simply concatenating histogram sequence estimated from different regions.