Digital Signal Processing Reference
In-Depth Information
If we try to use the likelihood of different values to optimize
the quantizer, we have to assume that the probability distribution
of the video data is known. Alternatively, the uniform quantizer
can be followed by some type of Huffman coding or differential
encoding, which will optimize the signal representation for the
minimum average number of bits.
Vector Quantization is also commonly used. In the preceding
discussion, a single set of values or signals is being quantized with
a given bit representation. However, a collection of related signals
can be quantized to a single representation using a given number
of bits.
For example, if we assume the pixel is represented in RGB
format, with 8 bits used for each color, the color red can be
represented in 2
8
or 256 different intensities.
Each pixel uses a total of 24 bits, for a total of 2
24
, or about
16 million possible values. Intuitively, this seems excessive as it is
unlikely that the human eye can distinguish between that many
colors. So a vector quantizer might map the 16 million possible
values into a color table of 256 total colors, allowing 256 combi-
nations of red, green and blue combinations. This mapping
results in just eight bits to present each pixel.
This seems reasonable, but the complexity lies in mapping
16 million possible inputs to the allowed 256 representations, or
color codewords. If done using a look-up table, memory of
16 million bytes would be required for this quantization, with
each memory location containing one of the 256 color code-
words. This is excessive, so some sort of mapping algorithm or
computation is required to map the 16 million possible color
combinations to the closest color codeword. It is hopefully
becoming clear that most methods to compress video, or other
data for that matter, come at the expense of increased complexity
and increased computational rates.
11.8 Decibels
Signal to Noise Power Ratio (SNR) is usually expressed in
decibels (denoted dB), using a logarithmic scale. The SNR of
a digitally represented signal can be determined by the following
equation:
¼
þ
1.76
Each additional bit of the signal representation gains 6 dB of
SNR. 8-bit representation is capable of a signal with an SNR of
about 48 dB, 12-bit can do better at 72 dB, and 16-bit will yield up
SNR
quantization
(dB)
6.02
(Number of bits)
