Adaptive multi-stream score fusion for illumination invariant face recognition

Quality variations of samples significantly affect the performance of biometric recognition systems. In case of face recognition systems, illumination degradation is the most common contributor of enormous intra-class variation. Wavelet transforms are very popular techniques for face or object recognition from images due to their illumination insensitiveness. However, low and high frequency subbands of wavelet transforms do not possess equal insensitiveness to different degree of illumination change. In this paper, we investigated the illumination insensitiveness of the subbands of Dual-Tree Complex Wavelet Transform (DTCWT) at different scales. Based on the investigations, a novel face recognition system has been proposed using weighted fusion of low and high frequency subbands that can adapt extensive illumination variations and produces high recognition rate even with a single sample. A novel fuzzy weighting scheme has been proposed to determine the adaptive weights during uncertain illuminations conditions. In addition, an adaptive normalization approach has been applied for illumination quality enhancement of the poor lit samples while retaining the quality of good samples. The performance of the proposed adaptive method has been evaluated on Extended Yale B and AR face databases. Experimental results exhibit significant performance improvement of the proposed adaptive face recognition approach over benchmark methods under extensive illumination change.