Lucid-GAN: An Adversarial Network for Enhanced Image Inpainting

Abstract

Image in-painting has been a significant research area which is being explored in the past decade. It has many diversified applications such as damaged image repairing and object removal. The hallmark algorithm for this application is EdgeConnect which comprises an edge generator and an image completion network. But, EdgeConnect is used to repair only damaged images. The process of repairing damaged images also faces the pitfall of enormous quality reduction after reconstruction. This paper presents a custom Generative Adversarial Network (GAN) - named Lucid-GAN on top of EdgeConnect to upscale the quality of the image. This decreases the noise in the reconstructed image, while parallely improving the SSIM, PSNR and the MAE scores by a considerable margin.