Quantitative evaluation of some edge-preserving noise-smoothing techniques

A quantitative evaluation of several edge-preserving noise-smoothing techniques is presented. All of the techniques evaluated are devised to preserve edge sharpness while achieving some degree of noise cleaning. They are based on local operations on neighboring points and all of them can be iterated. They are unweighted neighbor averaging (AVE), K-nearest neighbor averaging (KAVE), the edge and line weights method (EDLN), gradient inverse weighted smoothing (GRADIN), maximum homogeneity smoothing (MAXH), slope facet model smoothing (FACET), and median filtering (MEDIAN). The evaluation procedure involves two steps. First, the image is partitioned into regions based on the amount of spatial activity in a neighborhood of a pixel, where spatial activity is defined as local gradient. In the second part of the procedure an objective measure, the mean-square error, for each region of the partitioned image is obtained to evaluate the performance of the smoothing scheme at the corresponding level of spatial activity content. This evaluation procedure provides a convenient way to compare both the edge-preserving and noise-smoothing abilities of different schemes. The smoothing schemes were tested on a specially generated image with varying degrees of added noise and different edge slopes. The results of the comparison study are presented.