Fast Light-Weight Network for Extreme Image Inpainting Challenge

Abstract

Image inpainting has a wide range of applications. However, to this challenge existing inpainting models that usually have a large model size can hardly run fast, as memory and supported operations are much limited. In this paper, we propose a novel light-weight inpainting model in which we design three novel operations named Equilibrium Conv Mask-wise Gated Conv, Difference Conv and define a new loss function based on SN-patchGAN. In specific, the incorporation of Equilibrium Conv and Mask-wise Gated Conv not only reduces the model size and improve the efficiency, but also keeps comparative performance. For Difference Conv, it is benefit to handle big mask problem. Moreover, our proposed loss results in a better performance in recovering images with rich textures. Experimental results demonstrate our model is 1.43 $\times $ × speeding up and reduces the size by 2.37 $\times $ × compared with the state-of-the-art model.