Vector Quantization - Introduction to Data Compression - page 327

F I GU R E 10 . 26 Left: Reconstructed image using mean-removed vector quantiza-

tion and the Sena image as the training set. Right: LBG vector
quantization with the Sena image as the training set.

shown in Figure 10.26 . For comparison we have also included the reconstructed image from
Figure 10.16 . Notice the annoying blotches on the shoulder have disappeared. However, the
reconstructed image also suffers from more blockiness. The blockiness increases because
adding the mean back into each block accentuates the discontinuity at the block boundaries.
Each approach has its advantages and disadvantages. Which approachwe use in a particular
application depends very much on the application.

10.7.3 Classified Vector Quantization

We can sometimes divide the source output into separate classes with different spatial prop-
erties. In these cases, it can be very beneficial to design separate vector quantizers for the
different classes. This approach, referred to as classified vector quantization , is especially
useful in image compression, where edges and nonedge regions form two distinct classes. We
can separate the training set into vectors that contain edges and vectors that do not. A separate
vector quantizer can be developed for each class. During the encoding process, the vector is
first tested to see if it contains an edge. A simple way to do this is to check the variance of the
pixels in the vector. A large variance will indicate the presence of an edge. More sophisticated
techniques for edge detection can also be used. Once the vector is classified, the correspond-
ing codebook can be used to quantize the vector. The encoder transmits both the label for the
codebook used and the label for the vector in the codebook [ 157 ].
A slight variation of this strategy is to use different kinds of quantizers for the different
classes of vectors. For example, if certain classes of source outputs require quantization at a
higher rate than is possible using LBG vector quantizers, we can use lattice vector quantizers.
An example of this approach can be found in [ 158 ].