Multiscale morphology based illumination normalization with enhanced local textures for face recognition

A key challenge of face recognition is to obtain illumination invariant face images while preserving the discriminative features. The locations and shapes of small-scale features (e.g. eyebrows, eyes, nostrils, a mouth, etc.) are usually treated as key features for face recognition. However, it has also been observed that the local texture information of facial regions contains intrinsic facial features and needs to be enhanced to improve performance. To compensate for the illumination effects that appeared while extracting both the small-scale features and the texture information, we used multiscale morphological techniques. We used a generalized dynamic morphological quotient image (GDMQI) method based on Retinex theory and multiscale morphological closing to solve the artifact problem discussed in previous works. The proposed method consisted of two main steps: (i) illumination estimation and (ii) texture enhancement. The proposed method showed improved performance when using the CMU PIE, AR and Extended Yale-B databases.