Digital Signal Processing Reference
In-Depth Information
Table 11.2
Signal
Level
Closest 12 Bit
Representation
Hexadecimal
Value
Actual
Error
Error As A Percent
Of Signal Level
0.50300
0.502930
0 406
0.000070
0.0139%
0.50300
0.503418
0 407
0.000418
0.0831%
0.05030
0.050293
0 067
0.000007
0.0140%
0.05030
0.050781
0 068
0.000481
0.9568%
0.00503
0.004883
0 00A
0.000147
2.922%
0.00503
0.005371
0 00B
0.000341
6.779%
representation, the quantization noise is much less, allowing
even small signals to be represented with acceptable precision.
Another way of describing this is to introduce the concept of
signal to noise power ratio, or SNR. This describes the power of
the largest signal compared to the background noise. This can be
very easily seen on a frequency domain or spectral plot of a signal.
There can be many sources of noise, but, for now, we are only
considering the quantization noise introduced by the digital
representation of the video pixel values.
What we have just described is the uniform quantizer, where
all the step sizes are equal. However, there are alternate quan-
tizing mappings, where the step size varies across the signal
amplitude. For example, a quantizer could be designed to give the
same SNR across the signal range, requiring a small step size
when the signal amplitude is small, and a larger step size as the
signal increases in value. The idea is to provide a near constant
quantization error as a percentage of the signal value. This type of
quantizing is performed on the voice signals in the US telephone
system, known as
-law encoding.
Another possible quantization scheme could be to use a small
step size for regions of the signal where there is a high probability
of signal amplitude occurring, and a larger step for regions where
signal amplitude has less likelihood of occurring.
Uniform quantizing is by far the most commonly used scheme
in video signal processing, because we are often not encoding
simply amplitude, representing small or large signals. In many
cases, we could be encoding color information, where the values
map to various color intensities, rather than signal amplitude.
m
