Spherical harmonic based linear face de-lighting and compensation

摘要

Lighting causes great variance for the objects. That makes the robust face detection a challenging problem. Although the spherical harmonic method provides a good framework for object rendering and inverse rendering, the requirement for the surface normal makes its application in face detection not practical. We propose a planar lighting model for the face images and give out a linear method to estimate the incident or reflected light field of the image based on the spherical harmonic bases. Given the ratio of new light field to the original one, the image can be converted to any new lighting conditions according to the reflection theory. We propose a linear de-lighting and compensation (LDC) method to adjust images. First, we convert the illumination of the image to its inverse condition and then compensate it with the original image, where the weight of compensation is determined by the normalized light field of the image. In the face detection experiment, the LDC method shows its ability to facilitate the performance of face detector.