Image Processing Reference
In-Depth Information
Fig. 1.14
Generating a CA codebook via a learning mechanism and its use in a CA-VQ
compression scheme
As seen in Fig. 1.14, for a particular choice of the codebook
size
D
=
16
=
×
×
and block size
m
16 (4
4 window) the running of the 16
16 CA with
=
=
ID
4276938880 for
T
10 iterations would reveal the codebook. A zoom of both
×
the original image (512
512 pixels) and the recovered one using this codebook
for the most significant
4
bit-planes shows the typical losses of this compression
scheme. The PSNR is about 25 dB and the quality of the image is acceptable given
the rate of 1 bit per pixel. Further optimization of performance is still possible, for
instance the use of different, optimized codebooks, for each bit-plane, or a different
learning scheme.
In terms of computational complexity, the CA-VQ approach requires more com-
putational effort in the encoding stage and less effort in the decoding stage. It is an
opposite situation from the case of chaotic scan compression. For a message with
N
bits and a particular bit rate (codebook size
D
and block size
m
) assuming that
the codebook is calculated and stored, the compression process requires compar-
isons (between blocks and code-words in the codebook) being more effective for
large block sizes and small dictionaries (also ensuring highest compression but low-
est reconstruction quality). The above compares not so favorably to simply reading
K
N
samples in the chaotic scan method. But on the other hand, the complexity of
the CA-VQ compression process still remains linear in the number of pixels, being
much lower than for traditional compression methods. The decompression stage in
CA-VQ is
m
times faster (i.e.
N
<
/
m
operations of reading the codebook) and involves
Search WWH ::
Custom Search