Smoothing 2-D or 3-D Images Using Local Classification

We propose a new method for smoothing 2-D (or 3-D) images which preserves edge elements. Pixel data contained within a moving square (or cube), centered at each point under consideration, is tested to determine if it is homogeneous, or if a region boundary is present. If the area is homogeneous, the mean value of the whole area gives the smoothed value for the central point. If a region boundary is detected, then the data are classified into two clusters, and the smoothed value of the central point results from the mean value of the cluster including this point, if a context-based consistency criterion is respected. Two cases are considered: Gaussian noise and noise whose distribution is unknown. Results are given on a synthetic image corrupted by additive Gaussian noise, illustrating the efficiency of our method in relation with the signal to noise ratio. Results on a 2-D MR image are also shown