Impulsive Noise and Scratch Removal
with Preliminary Noise Detection

Noise becomes impulsive if most of the signal values change slightly, and at the same time some signal values change dramatically, in other words the change is clearly visible. Impulses may have different amplitude values, or all the same values. Noise commonly appears as black and/or white spots in images, i.e. the noisy pixels have either a very small or a very large value. This type of noise is often called salt-and-pepper noise. Pure salt-and-pepper is very easy to remove from images because the maximal value occurs rarely in actual images and thus just checking whether the pixel has a maximal or minimal value reveals if it is corrupted or not. A more realistic noise is modeled as bit errors in the signal values. Typical sources for this kind of noise are channel errors in communication or storage.

A different variant of impulsive noise is when the isolated scratches (usually truly white or truly black) appear on the image.

The classical approach to removing impulsive noise from images is to use the median filter or one of its modifications.

As a result disposing of noise leads also to a complete blurring of the image, i.e. it causes serious problems with details and edge preservation. The explanation of this fact is simple. Median filtering does not recognize if it is necessary to correct a current pixel or not and in general it corrects all the pixels.

To solve this general problem, it is necessary to detect the noisy pixels, to differentiate them from the other pixels and to correct only the detected noisy pixels.

To implement this detection, we propose the following solution.

It is known that the median filter is a sliding window filter, with the window N x N, where N is an odd number. So before filtering we create a local histogram in this window. As we know, impulsive noise is a “big” jump in brightness, so the brightness value of the central pixel belongs to one of the ends of the variation series created from the local histogram. We analyze where the value of central pixel of the window is positioned in the variation set. If it is positioned close to the ends, we can assume that it is an impulsive corruption and must be corrected.

This kind of analysis is used for the improvement of many filters: median, rank-order, cellular neural, etc. Nonlinear cellular neural filters for impulsive noise filtering are quite different from median filters. They are more selective in filtering impulsive noise. But in some cases they give too many false responses in detection and removal of impulses. As a result the filtered image may also be blurred or corrupted. The supplement of this filter by the described detector gives very good results.

So in general by implementing such preliminary noise detection, we achieve excellent results. For example, the median filter becomes gentler and less destructive to the image, but still very effective against impulsive noise and scratches.

Examples

(click on the images to enlarge,
make sure you don't have pop-ups disabled)

Scratch Removal

Scratch	Removed scratch