Applications of Emergent Phenomena - Systematic Design for Emergence in Cellular Nonlinear Networks - page 143

the EMTM built around a simple semi-totalistic CA can be used to map arbitrary
length signal sequences into easy to classify binary patterns.
8.4 Image Compression Using CA-Based Vector
Quantization
Image compression is a necessary ingredient whenever limited bandwidth chan-
nels are available. During the years several techniques and standards have been
developed [92, 93]. Traditional schemes of lossy image compression usually em-
ploy vector quantization (VQ) of image blocks (quite often 8×8 pixel blocks).
Discrete cosine transform (DCT) or other transforms (like Discrete Wavelet
Transform (DWT) or Integer Wavelet Transform (IWT)) are also used in com-
pression schemes. For instance, in [94] an optimized IWT transform architecture
is reported as having a “modest gate complexity” of about 56,000 gates (with an
area of about 3 mm 2 for a 0.35 Pm technology).
Both vector quantization (VQ) and transform coding are computationally inten-
sive algorithms, therefore an important research effort is directed to define various
circuit architectures requiring a minimum of hardware resources and power con-
sumption. Most of the computational complexity is explained by the need of complex
arithmetic operators (i.e. multipliers, adders, cosine evaluation etc.) to process the
pixels, as required by the traditional algorithms.
In [95, 96] a cellular automata based vector quantization (CA-VQ) method is
proposed for efficient image compression. The result is a radical simplification of
the circuitry, making it very attractive for integration with low power image sen-
sors. The method will be briefly exposed in the next. For more details the reader is
directed to [96].
Unlike traditional VQ systems where extensive learning on large datasets of
signals is employed, our codebooks are simply generated as emergent patterns in
cellular automata and therefore they are fully described with only a few bits of in-
formation (the ones needed to specify the ID and the initial state). The CA-VQ
method operates on binary images obtained as bit-planes of the original images
(e.g. the most-significant bits of all pixels are assembled into the most significant
bitplane) and therefore complex arithmetic circuits are eliminated. Only simple
circuits for Hamming distances and comparisons are needed in addition to memo-
ries storing the binary codebooks. The CA-VQ scheme outperforms the JPEG
standard for rates lower than 0.25 bits per pixel, while its simplicity makes it an
ideal candidate for integration in ubiquitous computing sensory systems.
8.4.1 Coding and Decoding Principle
As seen in Fig. 8.11 cellular automata are used to generate the codebook instead of
complex optimization algorithms operating on large training sets, as is the case in
the traditional vector quantization compression schemes.