On the selection of median structure for image filtering

Median filtering is a powerful tool for reducing noise in images, particular for long-tailed noise distributions. However, the choice of filter mask is critical. The proper choice depends on the image and noise statistics, which are often unknown. We propose cross-validation as a method for selecting a median filter structure directly from the corrupted image data. This method requires no knowledge of the statistics of the noise or image. We demonstrate the value of this method with several examples.<
>