基于最优传输模型的人脸颜色转换

现有人脸颜色转换算法仅基于图像的Lαβ颜色空间匹配均值和方差，因而仅限于线性变换且通常适用于自然图像；同时，现有算法利用人脸关键点信息来定位人脸五官的位置，但由此得到的人脸五官区域信息不是非常准确，通常需要进一步优化处理.针对上述问题，在自动获得人脸区域分割的基础上，为得到更自然的人脸颜色转换效果，基于最优传输模型，提出了一种新的人脸颜色转换的方法.首先利用全卷积网络自动得到人脸区域的分割信息，再利用最优传输模型获得对应的人脸区域的颜色转换结果.试验结果表明，所提算法在人脸五官分割的鲁棒性和人脸图像颜色转换的主观视觉上均得到明显的改善.
Abstract：Existing face color transfer algorithms are only based on Lαβ color space for matching mean and variance, and such transfer is limited to linear transformation and is usually applied to natural images. At the same time, these algorithms usually use facial landmark information to locate facial feature positions. However, the facial feature information thus obtained lacks accuracy and needs further optimization. To address these problems, a new face color transformation method was proposed based on the optimal transfer model to get the natural face color transfer effect. Firstly, the semantic information of the face was obtained directly by using the fully convolution network, and the corresponding face color transfer result was obtained by using the optimal transfer model. Experimental results show that the proposed algorithm is significantly improved in both the robustness of facial features segmentation and the subjective vision of face image color transfer.