Local selected features of dual-tree complex wavelet transform for single sample face recognition

In this study, a new method for single sample face recognition problem based on local dual-tree complex wavelet transform (DT-CWT) representation is proposed. The proposed system provides a number of countermeasures to neutralise the unwanted imaging conditions in face classification. First of all, two novel pre-processing steps, geometric and photometric normalisation are used to normalise the facial images. The employed geometric normalisation method is more consistent with the geometry and the shape of the face image, removing redundant information from classification. This step is followed by a new illumination normalisation approach in which useful information for the extraction of DT-CWT features is kept almost unaltered. Taking a local feature-based approach, DT-CWT features are then extracted regionally. By virtue of the effective normalisation stages and the employed multi-scale and multi-orientation DT-CWT features, the proposed locally selected DT-CWT method offers invariance to moderate real world image variations, such as illumination, expression, head pose, shift and in-plane rotation. The experimental evaluation of the method is performed on the widely used FERET and YALE databases, in an identification scenario, achieving promising performance.