Digital Signal Processing Reference
In-Depth Information
to be in the range of four to five ½ bits. Huffman coding, or similar
mapping of the values to bit codewords, can reduce the number
of bits used to transmit each color plane pixel to ~5 bits compared
to the original eight bits. This has been achieved without any loss
of video information, meaning the reconstructed video is iden-
tical to the original. This is known as lossless compression, as
there is no loss of information in the compression (encoding) and
decompression (decoding) process.
Much higher degrees of compression are possible if we are
willing to accept some level of video information loss, which can
result in video quality degradation. This is known as lossy
compression. Note that with lossy compression, each time the
video is compressed and decompressed, some information is lost,
and there will a resultant impact on video quality. The trick is to
achieve high levels of compression without noticeable video
degradation.
11.7 Quantization
Many of the techniques used in compression, such as the DCT,
Huffman coding and predictive coding, are fully reversible with
no loss in video quality. Quantization is often the principal area of
compression where information is irretrievably lost, and conse-
quently decoded video will suffer quality degradation. With care,
this degradation can be almost imperceptible to the viewer.
Quantization occurs when a signal with many or infinite
number of values must be mapped into a finite set of values. In
digital signal processing, signals are presented in binary numbers,
with 2
n
possible values mapping into an n bit representation.
For example, suppose we want to present the range
1to
þ
1
1) using an 8-bit fixed point number. With 2
n
or 256
possible values to map to across this range, the step size is 1 / 128,
which is 0.0078125. Let's say the signal has anactual valueof 0.5030.
Howclosely can this value be represented?What if the signal is 1/10
the level of the first sample, or 0.0503? And again, consider a signal
with value 1 / 10 the level as the second sample, at 0.00503. Below is
a table showing the closest representation just above and below
each of these signal levels, and the resultant error in the 8-bit
representation of the actual signal to sampled signal value.
The actual error level remains more or less in the same range
over the different signal ranges. This error level will fluctuate,
depending upon the exact signal value, but with our 8-bit signed
example will always be less than 1 / 128, or 0.0087125. This
fluctuating error signal will be seen as a form of noise (called
quantization noise), or unwanted signal in the digital video
(well, almost
þ
