高精度连续人脸轮廓提取

Refining Sparse Landmarks to Continuous and High-Quality Face Contour

提取连续且高精度的人脸轮廓线是很多图像应用重要的基础步骤.然而很多现有的轮廓提取方法并不能很好地应用在人脸轮廓上.为此提出一种快速可靠的人脸轮廓提取方法,能够在关键点提供初始化后提取到高精度连续人脸轮廓线.其主要步骤是先拟合关键点形成一条初始化曲线,沿其密集采样重叠的矩形区域,将整个人脸轮廓区域划分成很多小的区域;然后在每个局部的矩形区域提取出一条抛物线引导基于梯度的局部人脸轮廓线;最后从很多局部人脸轮廓线中,通过全局融合找到最终的人脸轮廓线.这种交叉验证的机制保证了最后结果的正确性.最后在LFPW和HELEN人脸数据集上进行了实验,结果表明文中方法能有效地提高人脸轮廓提取的精度.

Extracting continuous and high-quality face contour is an important preprocessing step of many high-level image applications. However, existing approaches are not dedicated algorithms for face contour extraction. In this paper, we find that it is feasible and economical to refine sparse landmarks to obtain a continuous and high-quality face contour, based on the fact that the landmarks can provide us with an initial face curve. The method proposed is composed of three steps: Firstly, e sample small and overlapped squares along the initial curve to cut the target face contour into small segments. We assume that any local segment of a face contour resembles a parabola, which is more reasonable than assuming the whole face contour as a parabola. Then locally in each square, we propose a parabola-constrained seam cutting algorithm to directly identify the face contour segment in the square. Finally, we propose a global seam integrating algorithm that extracts the target face contour from all the local segments. This cross-validation mechanism guarantees the success of the extraction of the final face contour. Experiments on two modern face datasets, i.e., LFPW and HELEN, demonstrate that our method significantly improves the results of the state-of-the-art face alignment approaches.