Learning filter selection policies for interpretable image denoising in parametrised action space

Abstract

AbstractThe denoising of images is an important research direction in computer vision. We consider the image denoising task as an estimation problem of the filtering policy related to image features, which is different from end‐to‐end image mapping. Commonly used simple filters such as gaussian filtering and bilateral filtering have fixed global denoising policies. However, the denoising policies of different filters can only adapt to limited image features. To solve this problem, we propose a method that applies different filters to different spatial ranges and adjusts the parameters of these filters simultaneously. Since not all filters can be easily transformed into differentiable forms and it is difficult to obtain paired datasets of filter action areas, we use reinforcement learning (RL) methods to estimate the spatial domain action range and adjustable parameters of filters, respectively. Furthermore, for removing higher intensity noise, simple filters can iteratively approximate higher‐order denoising policies and obtain more accurate and stable denoising results with the increase of iteration steps. Experimental results show that our proposed method can not only generate intuitive and interpretable denoising policies but also achieve comparable or better visual effects and computational efficiency than baseline methods.