Illumination Normalization and Skin Color Validation for Robust Face Detection

Sanghun Lee, Chulhee Lee

Abstract: The conventional Viola-Jones face detector may fail to detect faces under severe illumination conditions. The proposed method is based on the difference of Gaussian (DoG) that has been widely used to compensate for illumination effects. In the proposed method, we combine an original image and its DoG-filtered image, as a linear combination. This operation removed some illumination effects and improved face detection performance. Using the YCbCr color space, a skin color validation procedure was applied after face candidates were obtained using the proposed detector. Experiments using the Bao database showed that the proposed methods reduced over 50% of false positives.