Adaptive Varying Window Size Filtering of Color Images Based on Intersection of Confidence Intervals Rule

摘要

A class of linearonlinear filters with varying adaptive window size is studied. After some (optional) color transformation, for each of three color channels the window size of the applied mask-filters is considered as a parameter. The intersection of confidence intervals (ICI) rule is used for selection of the adaptive window size based on the filter's outputs obtained for different window sizes. In parallel five filters with symmetric and four quadrants masks are used. The ICI rule gives the adaptive window sizes for each of these filters and in a point-wise manner for each pixel of the image. This adaptive window size filters are able to suppress the noise efficiently provided that color edges are well preserved. The final filtering output is obtained by combining outputs of the mentioned five partial filters, each with the varying adaptive window sizes. This operation is produced for each color channel. Finally, we convert the estimates of the color image components back to RGB image. Originally, the ICI rule has been proved by theoretical and empirical studies to be efficient for linear and median filters. We show how this ICI rule can be modified and applied for color image filtering. Simulation experiments confirm that the ICI rule used for window size selection of the mean and median filters of the multichannel combined filters is able to significantly improve quality of color image filtering. The performance of the filters is characterized both by the accuracy and human visual perception criteria.