Image Inpainting via Enhanced Generative Adversarial Network

In this paper, we propose an Enhanced Generative Model for Image Inpainting (EGMII). Unlike most state-of-the-art algorithms using extra constraints to enforce the generator network recover semantic texture details, we construct an end-to-end network to generate both the image contents and high-frequency details progressively. Our model contains a two-phase generator and two discriminators. In our generator, the previous phase restores the structure information via convolutional encoder-decoder architecture, and the following phase captures high-frequency details via residual learning. For each generator phase, we define different objective functions and further optimize the entire network via a feed-forward manner. Moreover, for the generator in the second phase, we adopt a deep residual architecture, which also can eliminate the perceptual discontinuity on the border of the missing region. Experimental results on several public datasets demonstrate qualitatively and quantitatively that our model performs better than the state-of-the-art algorithms and can generate both realistic image contents and high-frequency details. Our code will be released soon.