基于生成对抗网络的人脸图像修复方法研究

Face Image Restoration Method Based on Generative Adversarial Networks

传统的图像修复方法在修复含有较大面积破损的图像时,不仅会出现明显的修复边界,而且与原图存在较大差异。为有效解决以上问题,提出一种基于生成对抗网络的图像修复方法。利用自编码器作为生成器网络结构,可以直接处理含有破损的人脸图像,然后引入LSGAN替换原始的生成对抗损失函数,解决网络训练过程中模型崩溃问题,最后结合基于距离的加权重建损失训练图像修复模型。在CelebA人脸数据集上的试验结果表明,本文方法所得到的修复图像较好地去除了修复边界,且人像五官细节更加贴合原始图像。

When the traditional image restoration method is used to repair the image containing a large area of damage,it will not only appear obvious restoration boundary,but also have great difference from the original image.In order to solve the above problems effectively,an image restoration method based on generative adversarial network is proposed.Using the autoencoder as the generator network structure,the face images containing damage can be directly processed.Then,LSGaN is introduced to replace the original generation adverse loss function to solve the problem of model collapse in the network training.Finally,the weighted reconstruction loss training model based on distance is combined with the image repair model.The experimental results on Celeba face dataset show that the restored image obtained by the method can remove the restoration boundary well,and the facial features details are more consistent with the original image.