Image denoising by adaptive kernel regression

This paper introduces an extremely robust adaptive denoising filter in the spatial domain. The filter is based on non-parametric statistical estimation methods, and in particular generalizes an adaptive method proposed earlier by Fukunaga [1]. To denoise a pixel, the proposed filter computes a locally adaptive set of weights and window sizes, which can be proven to be optimal in the context of non-parametric estimation using kernels. While we do not report analytical results on the statistical efficiency of the proposed method in this paper, we will discuss its derivation, and experimentally demonstrate its effectiveness against competing techniques at low SNR and on real noisy data.