分段仿射变换下基于泊松融合的正面人脸合成

针对目前正面人脸合成算法运算量大或合成图像失真较大的问题，提出一种基于分段仿射变换和泊松融合的正面人脸图像合成算法，将多幅输入图像用分段仿射变换（Piecewise Affine Warp，PAW）映射到正面人脸模板，并根据映射时产生的非刚性形变求得其对应的权重矩阵，进而获取每幅映射图像对应的变形掩膜，依次以这些映射图像为前景图像，以其对应的变形掩膜为泊松掩膜，并以上一次的融合图像为背景图像进行泊松融合，生成一幅平滑自然的正面人脸图像。实验结果表明，相比现有算法，该算法生成的正面人脸图像更加逼近真实正面人脸图像，而且很好地保留了输入人脸的个体信息。

Frontal face synthesizing based on Poisson image fusion under piecewise affine warp

In this paper, a frontal face synthesizing strategy based on Poisson image fusion and Piecewise Affine Warp（PAW） is proposed to solve the problem of large-scale computation cost or transformation distortion in general synthesizing methods. The multiple non-frontal input images are mapped to the frontal face template with PAW. The corresponding weight matrices are calculated according to the magnitude of deformation which can be used to obtain the foreground mask for Poisson fusion. Iterative fusion strategy is designed to synthesize one frontal image from multiple non-frontal images. In each step, the PAW image is used as foreground image. The deformation mask is used as foreground mask, and the fusion image of the previous step is used as background. Experiments show that the synthesized frontal image can perfectly preserve personal facial details and outperforms others in both subjective and objective evaluations.