Multi-scale Generative Model for Image Completion

Inpainting a large hole of an image is always a challenging task, especially for high-resolution images. Recently, deep learning based approach has brought excellent ability to deal with this issue. However, these methods always generate blurred edge and distorted details, which makes the inpainting images unreal. Besides, assessment for generated images is also challenging. Traditional assessment way uses various formulations, which only shows the partial image condition or distorted degree compared to original images, so this method cannot correctly reflect people's perception. To deal with these problems, we propose a multi-scale generative model which can gradually generate novel text to avoid distorted details, and the multi-scale losses can also eliminate blurred edges between inpainting results and the original region. To evaluate images, we also propose a deep convolutional neural network to do image quality assessment, which is closer to human perception. Experiments on multiple datasets show that our approach can generate more plausible inpainting results both on traditional evaluation criteria and our image quality assessment network.