Inpainting by deep autoencoders using an advisor network

Abstract

In this study, we tried to find a solution for inpainting problem using deep convolutional autoencoders. A new training approach has been proposed as an alternative to the Generative Adversarial Networks. The neural network that designed for inpainting takes an image, which the certain part of its center is extracted, as an input then it attempts to fill the blank region. During the training phase, a distinct deep convolutional neural network is used and it is called Advisor Network. We show that the features extracted from intermediate layers of the Advisor Network, which is trained on a different dataset for classification, improves the performance of the autoencoder.