Improved Nonlocal Means Denoising for Images with Tone Gradients

Image denoising is an essential feature of modern Image Signal Processor (ISP). As a part of mobile system, acceptability of denoising methods is determined by electrical power consumption, silicon footprint and architecture of ISP. Non-local means (NLM) denoising algorithm proposed by Buades et al. [1] has very convenient architecture to be implemented either as SoC or CPU/GPU based ISP. Unfortunately, it is prone to produce "staircasing effect" [2] on images having nonzero tone gradients. In this paper we propose a simple yet efficient extension of the original NLM algorithm that addresses the staircasing problem. Unlike existing solutions such as [3] or [5] we intentionally avoid kernel regression methods or multiresolution decompositions and focus on improvement of pattern matching process. As a result, our algorithm does not use costly regression computations and provides considerable subjective and objective performance gain (improvement in PSNR similar to 0.3dB). Also, the ideas explained here can be used in other image processing methods where block matching plays key role.