ABSTRACT
Nonlocal means (NLM) is one of the classic patch-based methods for random noise attenuation. It assumes that a significant amount of redundant information exists in similar patches, which can be used to restore the original data. However, this method is computationally expensive due to a large number of overlapping patches. In addition, because this method uses a weighted average of patches to suppress noise, when applied to the data with complicated structures, the “average effect” may appear in the denoised results. We have implemented the NLM in the frequency-space domain, which can be called adaptive frequency-domain nonlocal means. This novel strategy will significantly reduce the computational cost and improve the quality of the final result compared to traditional NLM. Considering the impact of filtering parameters on the final result, we also build a mapping relationship between the noise strength and the filtering parameters, which could help obtain the denoised result with less signal leakage. We provide synthetic and field examples to show the superiority of this method over the traditional f-x prediction and NLM methods.