Inspired by the edge-preserving filter used in computational photography, we propose a method of illumination normalization for face recognition under varying lighting conditions. The brightness layer of a face image is accurately decomposed into a reflectance layer and a shading layer with a weighted least square filter. The histogram of the reflectance layer is remapped to that of sample images selected based on a visually optimal criterion to obtain an image with the best visual quality. Experimental results on the YALE-B and CMU-PIE face databases demonstrate effectiveness of the proposed method under varying illumination conditions.