Visual quantization

In gray scale quantization of monochrome images, if the number of quantization levels is not sufficient, a phenomenon called "contouring" become visible. When groups of neighbouring pixels are quantized to the same value, regions of constant gray levels are formed, whose boundaries are called "contours". Uniform quantization of common images where pixels represent luminance, require 8 bits (i.e. 256 gray levels). Contouring effects start becoming visible at or below 6 bits/pixel. A mean square quantizer (MSQ) matched to the histogram of the image may need only 5 to 6 bits/pixel without any visible contours. As histogram of images vary drastically optimum quantization with 8 bits/pixel is usually used. In evaluating quantized images, human eye is quite sensitive to contours and errors that affect the local structure. Note that contours do not contribute very much to mean square error. Thus a visual quantization scheme is used to hold quantization contour below the level of visibility over the range of luminances to be displayed.

Methods of visual quantization:

Contrast quantization

Since visual sensitivity is nearly uniform to just noticeable changes in contrast, it is more appropriate to quantize the contrast function.

Two non linear transformations used to represent contrast are, ,

or

and

where u is the luminance; are constants chosen as and are constants.

For the given contrast representation, we simply use the MMSE quantizer for the contrast field.

To display the image, the quantized contrast is transformed back to luminance value by inverse transformation.

Experimental studies indicate that a 2% change in contrast is just noticeable.

i.e

This needs 50 levels in a scale of contrast from 0 to 1.Therefore, if uniformly quantized, the contrast scale needs 50 levels or 6 bits.

However with optimum MS Quantizers 4 to 5 bits could be sufficient.

Pseudo Noise quantizing

It is a method of suppressing contouring effects. This is done by adding a small amount of uniformly distributed pseudo random noise to the luminance samples before quantization. This pseudo random noise is called dither. To display the image, the same (or another) pseudo noise is subtracted from quantizer output as shown in Figure (3.6.8) above. The effect is that in the regions of low luminance gradients (which are regions of contours) the input noise causes pixels to go above or below the original decision level, thereby breaking the contours. However, the average value of quantized pixels is about the same with and without the additive noise. During display, the noise tends to fill in the regions of contours in a such a way that the spatial average is unchanged. The amount of dither added should be kept small enough to maintain the spatial resolution but large enough to allow the luminance values to vary randomly about the quantizer decision levels. The noise should usually affect the least significant bit of quantizer. Reasonable image quality is acheivable by a 3-bit quantizer.