MTPFK: Multi-scale Transformer Joint Predictive Filter Kernel for Image Inpainting

In the task of image inpainting, it is common to utilize a CNN-based encoder-decoder architecture to extract the feature information from the damaged image, achieving satisfactory restoration results. However, these methods often struggle to achieve high-quality restoration for images with varying degrees of damage. In this paper, propose a two-stage inpainting model. Firstly, leverage the powerful contextual capturing capabilities of the Transformer to form a coarse recovery network, so as to roughly fill holes of different sizes. Secondly, employ a predicted filtering kernel network to perform fine restoration, building upon the coarse restoration. Method conducted qualitative and quantitative experiments on the CelebA and Places2 datasets, demonstrating the superiority of our proposed method.