The Image Processing Corner

Several situations of noisy image exist.

The first one is the case of a video. The main application is old film restoration. In this case, common information between successive images may be exploited. This is especially true for fix scenes.

The other situation is a single image. A typical example is the digital photo taken at the limit of the sensor's sensitivity (too dark scene).

Many image noise sources may be found (non exhaustive list):

• Gaussian noise. A gaussian random variable has been added by the system to the "ideal" image. In the previous case of the digital photo, the source is the precision of the CCD or CMOS sensor emphasized by a high gain.

• Uniform noise. A uniform random variable has been added by the system to the "ideal" image. It is a lab case, not realist but simple. Visually, it is close to gaussian noise (only the distribution changes).

• "Salt and pepper" noise. Some pixels are saturated. Visually, some black and some white pixels appear in the image, with a given density. The most usual sources are dead pixels in the sensor, due to its manufacture process (same issue with LCD monitors). Cameras "hide" dead pixels by detecting them at manufacture stage and by using a closest neighbor replacement mechanism when taking a photo.

• Periodic noise. A periodic noise has been added by the system to the "ideal" image. The source may be due to a scanning process, or a JPEG compression (8 pixel wide blocks).

The different noises are illustrated below. The first row gives the noisy images and the second row shows a detail with colors which enhance contrast.

Next pages show some methods to reduce the noise.