Image artefact removal using deep learning

Abstract

Compression artefacts refer to the unwanted and unpleasant distortions which contaminate an image when it has been compressed. The removal of these artefacts is an essential task since they affect the usability, impact or informativeness of an image. There are multiple techniques for image artefact removal. There are more traditional methods such as image adaptive filtering, bilateral filtering, methods using POCS and the SSRQC method. There are also newer approaches using deep learning such as AR CNN, DnCNN and GAN. This report implements variations of the deep learning methods, namely a combination of the AR CNN and DnCNN in the form of a Residual AR CNN, a GAN which uses PatchGAN for its discriminator and a Residual GAN which uses residual learning. These methods are tested on images with varying degrees of compression and evaluated. The results are then compared to non-deep learning approaches, SSRQC method and Bilateral filtering, in order to see the differences in performance