Databases Reference
In-Depth Information
Quantizer
output
215
215
190
190
165
165
140
Weight (lb)
140
115
115
90
90
65
65
40
40
50 60
Height-weight quantizer
70
Height (in)
Weight quantizer
40
50
60
Height quantizer
70
F I GU R E 10 . 2
The height/weight scalar quantizers when viewed in two
dimensions.
Note that we have not said how we would obtain the locations of the quantizer outputs
shown in Figure
10.3
. These output points make up the codebook of the vector quantizer, and
we will be looking at codebook design in some detail later in this chapter.
We can see from this example that, as in lossless compression, looking at longer sequences
of inputs brings out the structure in the source output. This structure can then be used to
provide more efficient representations.
We can easily see how structure in the form of correlation between source outputs can
make it more efficient to look at sequences of source outputs rather than looking at each sample
separately. However, the vector quantizer is also more efficient than the scalar quantizer when
the source output values are not correlated. The reason for this is actually quite simple. As
we look at longer and longer sequences of source outputs, we are afforded more flexibility in
terms of our design. This flexibility in turn allows us to match the design of the quantizer to
the source characteristics. Consider the following example.
