A robust improvement of the ICI rule for signal denoising

Denoising is an important issue in signal processing. Noise, caused by sensors or by quantization effects during digitalization or compression, can significantly influence the processing results. Hence, removing the noise, or extracting the signal with minimal distortion is a valuable objective for many applications. In this paper, we propose a novel robust denoising method. It originates from a local polynomial approximation scheme, where the estimation window is being fit to signal properties using the intersection of confidence intervals rule, or an improved relative intersection of confidence intervals method. Under a very weak set of assumptions (symmetry of the noise distribution), we modified the preceding methods into a robust median based approach. One of the two proposed methods significantly overperformed the competitive methods, in terms of lower sensitivity to predefined parameters, noise distribution and presence of outliers.