Progressive Inpainting Strategy with Partial Convolutions Generative Networks (PPCGN)

Abstract

Recently, there have been great advances in many one-stage image inpainting methods. They may have a slight advantage in computation time but lack sufficient context information for inpainting. These inpainting approaches can not inpaint large holes naturalist. This paper proposes a progressive image inpainting algorithm with partial convolution generative networks for solving the above problem. It consists of a generator with partial convolution layers, a fully convolutional discriminative network, and a long short-term memory (LSTM) module. PPCGN has four steps to inpaint the image. Each Step will concentrate on a specific area for inpainting. The final generation results are completed by the cooperation of these four steps, which are connected through the LSTM module. Due to the partial convolution module and LSTM structure characteristics, our method has a good advantage in restoring the images with large holes and achieves better objective results, increasing 1.46 dB and 0.44 dB on the Paris Street View dataset, and CelebA dataset, respectively.