基于混合能量活动轮廓模型的人脸分割方法

Face Segmentation Based on a Hybrid Energy Based Active Contour Model

由于受到面部五官、饰物等因素的影响,传统几何活动轮廓模型获取人脸外轮廓会产生凹陷、分片等现象.针对人脸图像的特点,将边缘外张力能量及肤色能量与全局能量结合,提出一种基于混合能量泛函的几何活动轮廓模型,有效地避免了这些问题.首先,根据演化曲线的邻域信息赋予边缘点向外的张力,使曲线能够克服面部特征及面部饰物的干扰,引导其向外轮廓方向演化.鉴于肤色是面部最重要的特征,提出肤色能量,进一步提高了模型的鲁棒性.此外,提出一种基于单高斯模型的改进算法,能够估计出接近实际人脸外轮廓的初始位置,为轮廓演化奠定了基础.在两个公共人脸库上进行测试,该方法能够得到准确的人脸分割效果;以手工分割的结果为基准,该算法定位精度明显优于传统的全局能量模型和局部能量模型.还用日常照片创建一个包含不同姿态、光照、复杂背景等因素、复杂的人脸库,分割结果表明,该方法能够克服这些因素的影响,取得了准确而稳定的人脸分割结果.

Influenced by factors like facial features, accessories, facial outer contours are extracted by the traditional geometric active contour models and conatin depressions and result in fragmentation, etc. To address these problems, according to the characteristics of human face image, the study proposes a hybrid energy based geometric active contour model via combining the energies of contour outer tension force and skin color with the global energy. First an outwards tension force, computed by neighborhoods of contour points, is added to the contour. This force makes the curve insusceptible to the facial features and accessories, but move towards to the facial outer contour. As skin color is the major feature of a human face. Skin color energy is integrated to ensure a more robust algorithm. Finally, an improved skin tone detection model is proposed based on the single Gaussian function. It could generate initial position that are close to the real facial contour, laying a good foundation for contour evolution. The proposed method gives essentially accurate face segmentations on two public face databases. Take the manually segmentations as the ground truth, the proposed method compares favorably to both traditional global and local energy algorithms. Next a more challenging set containing 100 faces of life photos with variances in pose is introduced with illumination and backgrounds. Segmentation results have validated that the proposed method could extract outer facial contour steadily and accurately under such variances.