Robust extensions to guided image filtering

Abstract

Image denoising is commonly regarded as a problem of fundamental importance in imaging sciences. The last few decades have witnessed the advent of a wide spectrum of denoising algorithms, capable of dealing with noises and images of various types and statistical natures. It is usually the case, however, that the effectiveness of a given denoising procedure and the complexity of its numerical implementation increase pro rata, which is often the reason why more advanced solutions are avoided in situations when data images have relatively large sizes and/or acquired at high frame rates. As a result, substantial efforts have been recently extended to develop efficient means of image denoising, the computational complexity of which would be comparable to that of standard linear filtering. One of such solutions is Guided Image Filtering (GIF) - a recently proposed denoising technique, which combines outstanding performance characteristics with real-time implementability. Unfortunately, the standard implementation of GIF is known to perform poorly in situations when noise statistics deviate from that of additive Gaussian noise. To overcome this deficiently, in this note, we propose a number of modifications to the filter, which allow it to achieve stable and accurate results in the case of impulse and Poisson noises.