Selective denoising in document images using reinforcement learning

Abstract

Image denoising deals with removal of unwanted noise from images. While there have been many techniques that can be applied to denoise a given input noisy image, the methods process an image in its entirety, assuming that the noise uniformly affects the entire image. For inputs where the noise affects a localised part of the image, applying methods that attempt to denoise the entire image can adversely affect the clean portions. To address this problem, we propose a deep reinforcement learning-based framework aiming to overcome this limitation and achieve better results for images with non-uniformly distributed noise. We propose a two-step procedure that first identifies the noisy patch and then denoises the extracted patch. We use a reinforcement learning-based approach for noise localization and use PixelRL for noise removal. We have prepared a comprehensive dataset specifically for the noise localization problem, and noise patches are induced in clean document images using various noise patterns, such as Gaussian noise, coffee stains, and ink bleeds.