Multi-features fusion based CRFs for face segmentation

Face segmentation is quite challenging due to the diversity of hair styles, head poses, clothing, occlusions, and other phenomena. To improve the accuracy of face segmentation from the images with complex scenes, we present a method based on Conditional Random Fields (CRFs) in this paper. The CRFs model is defined on a graph, in which each node corresponds to a superpixel and each edge connects a pair of neighboring superpixels. The features of color and texture are used to define the node energy function, and the position distance and differences of features between adjacent superpixels are used to define the edge energy function. Segmentation is performed by inferring the CRFs model built by fusing node energy function and edge energy function. We evaluate the performance of the proposed method on two unconstrained face databases. Experimental results demonstrate that the proposed method can efficiently partition face images into regions of face, hair, and background.