A multi-scale context encoder for high quality restoration of facial images

In recent years, with the growth of data, storage and computing power many challenging problems can be solved. One of the most challenging and interesting tasks is to recover data in pixel level of image without using any additional information. Context encoder, a combination of auto-encoder and GAN, uses an unsupervised visual feature learning algorithm to predict missing pixel data in images. This paper presents a novel method to extract features in context encoder to restore high quality facial image. Instead of computing features from a single scale input image, the proposed method utilizes features from multiscale input images to create features in the encoder step. The features are then fed into the decoder to create image with the needed information after feature aggregation in different ways. The restored images after using decoder and the original ones are then provided to a discriminator to make the network creating a looked-real image. Experimental results on some public data sets such as CELEBA-HQ and LFW show that the proposed method with features in multi-scale achieves a promising result compared with the original one.