An Empirical Evaluation of Denoising Techniques for Streaming Data
Jeremy Thompson · 2014
We investigate denoising techniques for streaming data.This is data that is analyzed while it is collected.We consider spatial filters, such as the box filter, Gaussian smoothing, and the bilateral filter; frequency-based techniques, such as fast Fourier transform and wavelet transform, combined with thresholding of the coefficients; and a statistical neighborhood filter, the nonlocal means algorithm.We discuss practical concerns for incremental implementation, such as edge treatment, incremental updating, and parameter stability.These methods are applied to both synthetic data and real world data.Based on several carefully designed experiments, we note situations when these methods could fail and make recommendations for their use with real world data.Specifically, we suggest the use of the bilateral filter alone or a combination of the bilateral filter and non-local means algorithm.