关键词: 图像修复, 深度学习, 生成式对抗网络, 双编码器, 跳跃连接

Abstract:

In order to solve the problem of information loss in the damaged area during image restoration and to repair any damaged area in the image, a dual-encoder model is designed to independently encode the mask and image, reconstruct the image using the mask features, and reduce the mask information loss, adding skip connections to supplement image information loss due to down sampling and accelerate network convergence, introducing adversarial training to improve the quality of reconstructed images. The training and testing results on the places2 dataset show that the image restoration effect of this method has good performance in accuracy and globality, and can be used for image restoration of various types of masks.

Key words: image restoration, deep learning, Generative Adversarial Network（GAN）, dual-encoder, skip connections