Abstract: In order to repair facial images with large areas damaged efficiently, a method is introduced that using encoder-decoder structured Convolutional Neural Network（CNN） as the generative model, and skip-connections are added between partial layers of it to enhance its structured prediction power, at the same time, an anti-training strategy is introduced to optimize the generative model. This model first trains the discriminator to identify the real image, and then uses the discriminator to judge whether the generator’s output which obtained after the damaged image is input to the generator is true or faker. The discriminator is used to provide gradient to the generator to optimize the generator. Combining the structured prediction power of CNN and the optimization ability of GANs’ antagonism strategy, it improves the results of image completion obviously. The experimental results on the CelebA face dataset show that the proposed method is superior to other methods in repairing images with large areas of damage.

Key words: deep learning, Generative Adversarial Networks（GANs）, Convolutional Neural Network（CNN）, image completion, skip-connection

摘要： 为了有效地修复大面积破损的面部图像，使用了解码器-编码器结构的卷积神经网络作为生成模型，并在其部分层之间增加skip-connection，以增强生成模型的结构信息预测能力，同时引入对抗训练策略优化生成模型。该模型首先训练一个判别模型识别真实图像，再利用其判别待修复图像输入生成模型后所得到的输出是否为真实，以此为生成模型提供优化梯度。结合了卷积神经网络的结构信息预测能力和GANs对抗策略的优化能力，提高了图像补全的效果。在CelebA人脸数据集上进行的实验结果表明，该方法在补全大面积破损的图像任务上性能明显优于其他方法。

关键词: 深度学习, 生成对抗网络（GANs）, 卷积神经网络（CNN）, 图像补全, 跳跃连接