A robust framework for face contour detection from clutter background

Xun Gong, Xinxin Li, Lin Feng, Ran Xia

Abstract: When applying Chan–Vese (C–V) model to segment a face from an image, the result is always influenced by the initial position, especially in a real scenarios with clutter background. An improved skin tone detection model is proposed based on the Gaussian function, which could generate an accurate initial contour for C–V model. As the single Gaussian model (SGM) only uses the prior information to determine the likelihood of a pixel, it lacks adaptability for an image with lighting and poses variances. A more adaptive SGM (ASGM) is proposed in this paper that, by updating the model parameters according to the input image, could provide a more stable approximation of face region. And then, we apply the estimated initial contour to C–V model for precise face segmentation. Tests conducted on a public face dataset (220 images with pose and illumination changes) have shown that the accuracy of skin region detected by ASGM is at a ratio higher than 99%, which lays a good foundation for further segmentation by the C–V model. Extensive experiments have validated effectiveness and efficiency of our system in face segmentation.