IMAGE INPAINTING FOR ARBITRARY HOLES USING CUSTUMIZED RESIDUAL BLOCK ARCHITECTURE WITH PARTIAL CONVOLUTIONS

Abstract

Recently, learning-based algorithms for image inpainting achieve remarkable progress dealing with squared or regular holes. However, they still fail to generate plausible textures inside damaged area because there lacks surrounding information. In this paper, motivated by the residual learning algorithm which aims to learn the missing information in corrupted regions, thus facilitating feature integration and texture prediction we propose Residual Partial Convolution network (RBPConv) based on encoder and decoder U-net architecture to maintain texture while filling not only regular regions but also random holes. Both qualitative and quantitative experimental demonstrate that our model can deal with the corrupted regions of arbitrary shapes and performs favorably against previous state-of-the-art methods.