基于生成对抗网络的人脸图像彩色化方法

Face Image Colorization Based on Generative Adversarial Networks

提出了一种新的基于生成对抗网络的人脸图像彩色化方法.所提出的网络结构包含两组生成对抗子网络,每个子网络由一个生成器和判别器组成.其中,一个对抗子网络A(包含生成器A和判别器A)实现从灰度图像到彩色图像的翻译过程,另一个子网络B(包含生成器B和判别器B)反转该过程,即生成器B对称地使用生成器A的最终输出图像作为输入,用来重建原始的人脸灰度图像.其中,网络中的循环损失进行图像重建,而生成损失和对抗损失用来保证生成的图像更加接近真实图像.实验结果表明,这种结构设计不仅能实现自然逼真的人脸图像彩色化,还能同时保证人脸的身份属性不变.

In this paper,a novel face image colorization method was proposed based on a generative adversarial network.Two groups of generative adversarial sub-networks were involved in the network structure,a generator and a discriminator.One of the sub-networks A(containing generator A and discriminator A)could implement the translation process from gray-scale images to color images,while another sub-network B(containing generator B and discriminator B)could reverse the process.Taking the generated image of sub-networks A as input,the sub-networks B could reconstruct the original gray face image.The whole structure was arranged to ensure the invariance of human face identity,the loop loss in the network was for image reconstruction,the generative loss and adversarial loss were used to make the generated image close to the real image.The experimental results show that this structure can not only achieve a natural and realistic face image,but also ensure that the identity of the face.