readme2

Vector Quantization压缩算法

EXAMPLE:
Three training images: man, couple, crowd which are 512X512 and 8bpp.
One test image: lena which is also 512X512 and 8bpp.  We want to
design a tree-structured VQ of dimension 4 made from 2X2 blocks of the
images with average rate 5 bits per vector.  The data is stored as
"unsigned char" data.  So check the tsvq.h file (see the first real
line of file) and make sure that DATATYPE is defined as:

#define DATATYPE	unsigned char

Note: If you are using 16bpp then define data as short int, etc.

Then if needed, compile the programs
make all
make clean

Put all of the training images into one file:

cp man training_data
cat couple >> training_data
cat crowd >> training_data

Block the training data into vectors taken from 2X2 chunks of the
image using the "block" program from the stdvq directory.  The
training_data now has three images, so there are 3*512=1536 rows and
512 columns:

block -i training_data -r 1536 -l 512 -h 2 -w 2 -o training_data.TS

The output is a file called training_data.TS where TS is for training
set.  Now the training data are ready to use.  Use the "tsvq" program
to make a tree-structured codebook with rate 5 bits per vector.
Recall that the vector dimension is 4.

tsvq -t training_data.TS -c codebook -d 4 -r 5.0
tsvq -t training_data.TS -c codebook -d 4 -r 5.0 -B

The first will output an unbalanced tree-structured codebook called
"codebook" with average rate 5 bits per vector.  It will also output a
file called "training_data.TS.stat" which contains the rate and
distortion and other information of the training data at each node of
the codebook.  The information is needed for pruning.  The second will
output a balanced tree-structured codebook called "codebook" with
fixed rate 5 bits per vector.  It will also output a file called
"training_data.TS.stat" which contains the rate and distortion and other
information of the training data at each node of the codebook.  The
information is needed for pruning.

To encode the test image, the test image must first be arranged into
vectors using the "block" program.  Then the "tsvqe" program is used
to encode the test image.  Finally, the "unblock" program from the
stdvq directory is used to arrange the encoded/decoded vectors into a
decoded image.  Note that the "tsvqe" program does both encoding and
decoding on the data and does not output the indices of the compressed
vectors.

block -i lena -r 512 -l 512 -h 2 -w 2 -o lena.TS
tsvqe -c codebook -i lena.TS -o lena.TS.encoded
unblock -i lena.TS.encoded -r 512 -l 512 -h 2 -w 2 -o lena.final

_________________________________________________________________________
EXAMPLE:
You could also use the "-R" option in "tsvqe" to output an image
called rate.dat which shows the instantaneous rate used at each pixel
so you can see where more bits are being used throughout the image.

tsvqe -c codebook -i lena.TS -o lena.TS.encoded -R
unblock -i rate.dat -r 512 -l 512 -h 2 -w 2 -o rate.date.final

_________________________________________________________________________
EXAMPLE:
Prune a tree-structured codebook with average rate 8 bits per vector
to yield a tree-structured codebook with average rate 5 bits per
vector.

First we create a large tree-structured codebook using "tsvq".  Note
that it doesn't matter if the tree is balanced or not. Use the
"-s" option to name the output statistics file.

tsvq -t training_data.TS -c codebook -d 4 -r 8.0 -s stats
tsvq -t training_data.TS -c codebook -d 4 -r 8.0 -s stats -B

Use the "prune" program to produce a list of all optimal sub-trees.
The "-E" option can be used to do entropy-constrained pruning if
desired.

prune -c codebook -s stats -o nested > list_of_subtrees
prune -c codebook -s stats -o nested -E > list_of_subtrees

The file "list_of_subtrees" is a text file which lists all of
the subtrees by number with their rates (or entropies) and distortions.
The file "nested" contains the nested subtree structure.
Look through the "list_of_subtrees" file to find the subtree with
the desired rate(entropy) - distortion.  Use the "select"
program to choose the subtree:

select -c codebook -s nested -n # -o codebook.#

where # is the number of the subtree selected in the "list_of_subtrees."
Then you are ready to encode a test image:

block -i lena -r 512 -l 512 -h 2 -w 2 -o lena.TS
tsvqe -c codebook.# -i lena.TS -o lena.TS.encoded
unblock -i lena.TS.encoded -r 512 -l 512 -h 2 -w 2 -o lena.final

______________________________________________________________________
EXAMPLE:
Note that another way to generate a "stats" file is to encode a
different set of training data using "tsvqe" with the "-D"

tsvqe -c codebook -i training_data2.TS -o junk -D
rm junk

This will output a file called training_set2.TS.stat for the codebook
which can then be used in the "prune" program.

prune -c codebook -s training_set2.TS.stat -o nested > list_of_subtrees