SVD-Structural Similarity in the Wavelet-Gabor Domain: Improved Confidence for Face Recognition under Noise, Blur and Haze

In this work we propose and investigate the performance of a new similarity measure based on Singular Value Decomposition (SVD) and structural similarity in the wavelet and Gabor domains. The reasoning behind this combination is to utilize SVD in getting independent components, wavelet decomposition to get the complex frequency features and Gabor filtering to get textural features. A comparison has been made versus correlative and structural similarity measures like SSIM (Structural Similarity Index Measure), Complex-Wavelet SSIM (CWSSIM) and FSIM (Feature-Based Similarity). In these tests, a reference image is tested for similarity against several face images in a database under adverse conditions like noise, blur and haze. A new haze formation approach has also been proposed. Similarity level and similarity confidence are taken as the performance measures. Two confidence measures, different in strength of confidence, have been proposed and tested versus a recently-proposed confidence measure that relies on the difference between the maximal similarity (best match in the database) minus the second maximum similarity (second-best match in the database). Simulation using AT&T database has shown that the proposed SVD-Structural Similarity in Wavelet-Gabor Domain (SVWG) outperforms existing measures by far. SVWG can give more robust decisions (near-optimal confidence); also, can work under more adverse conditions (lower SNR, more blur or haze) where other similarity measures fail.