Most expressive feature extracted by half-quadratic theory and multiresolution analysis in face recognition

In this paper, a new model for extracting most expressive feature by half-quadratic theory and multiresolution analysis is proposed. The new model has two main advantages over some famous feature extraction algorithms. First, it is robust towards a number of outliers, especially when applied in face recognition. A robust M-estimator has been introduced into the new model to estimate robust scale parameter, so the generalizing ability of the new model is enhanced. Second, the new model could extract features in multi-scale space with the help of 2D wavelet decomposition. The validity of the new algorithm is confirmed by applied in face recognition. Yale and FERET face database are used in experiments. The experimental results have confirmed that the new model could cope with various outliers in face recognition, such as occlusions, making up, lighting conditions and incomplete face images, etc. And it could outperform some famous algorithms (PCA, RPCA, FLD and ICA)