| markootje@gmail.com 2007-07-29, 6:55 pm |
| Hi,
I'm currently working on an assignment where I have to implement the
Tree Structured Vector Quantization on images. The software currently
only supports normal VQ.
I've got a question about the practical aspect of creating a codebook
for TSVQ. I've studied many papers, but none tells me how to
practically construct a codebook from a train image. Hope somebody
here can help me out!
I'm thinking about the following algorithm.
Let's say the train image is 256*256 pixels, the TSVQ size is 8, and
the codebook vector size is 4 * 4 pixels:
On level 1 there are 8 codebooks, each of size 32*32 (256/8=32).
On level 2 there are 64 codebooks, each of size 4*4 (32/8=4).
For each codebook the LBG algorithm is applied (so 8 times on the
first level, and 64 on the second level).
Is this assumption of how to implement the TSVQ right? If yes, then I
only should think up how to come to a codebook vector size like 4*2 or
2*1.
Hope somebody can help me!
Kind regards,
Marko
|