Generative Adversarial Networks for Facial Image Inpainting and Super-Resolution

Abstract

Inpainting helps to fill in the lost data in visual images. Inpainting techniques also refer to unusual image editing in distorted regions. These include areas that are noisy, blurred and watery areas. The most appropriate pixel values must be replaced in these regions to achieve good performance. Artists used to play it, and still now, pieces that are not in the picture can be inpainted in the same manner, though it takes more time. In the present age of automation, inpainting can be automated to obtain quicker and better outcomes by deep learning technologies. In this area, many of the latest techniques have been created, however, many methods produce blurred findings and data loss. Two adversarial networks are used to achieve this task, where first network aims at inpainting and the second network aims at super-resolution. The input generated as a part of first stage network is passed on to the second stage super-resolution network to overcome blurriness that is caused in the initial inpainting network. The network efficiency is determined in terms of increased PSNR obtained which is 28.19 dB with less training period of approximately 14 hours in comparison with other network models which performs similar task.