基于CycleGAN改进的异质人脸图像合成算法

An Improved Heterogeneous Face Image Synthesis Algorithm Based on CycleGAN

异质人脸图像合成是以一类人脸图像为输入,合成另一类在图像质量、风格等方面与输入图像不同的新人脸图像。异质人脸图像合成要求在合成前后保持身份属性不变,同时要求合成图像的质量、风格等与另一类尽可能相近。原有的一些异质合成方法在这两方面均有所欠缺,经过对异质人脸合成的研究,针对这两点,对经典图像翻译模型Cy⁃cleGAN进行改进,并在人脸数据集上定量实验,验证合成方法的有效性。

Heterogeneous face image synthesis takes one kind of face image as input and output another kind of new face image which is different from the input image in image quality and style.Heterogeneous face image synthesis requires that the identity attributes remain unchanged before and after synthesis,and the quality and style of the synthesized image should be as close as possible to another type.Some of the het⁃erogeneous synthesis methods are deficient in these two aspects.After the research on heterogeneous face synthesis,the classical image translation model CycleGAN is improved for these two points,and the effectiveness of the synthesis method is verified by quantitative exper⁃iment on face data set.