Semi-Totalistic CNN Genes for Compact Image Compression
Radu Dogaru, Ronald Tetzlaff, Manfred Glesner · 2006
It is shown that using several tools for detecting emergent computation, a series of several tenths of genes useful for a novel, compact image compression scheme, were identified within the space of all 1024 semi-totalistic cellular automata (CA) with 5 neighbors (von Neumann neighborhood). Such cellular automata can be easily implemented on the CNN-UM using "B"-templates with only 5 elements. Spatio-temporal binary patterns with a fractal characteristic emerge in CNNs using such genes. These patterns are then used as codebooks for a simple-to-implement vectorial quantization scheme called CNN-VQ. Gray level images are split into bitplanes and each 8times8 block is approximated with its closest (in terms of Hamming distance) code-word form the CNN-generated codebook. Decoding is straightforward and includes a median filter to remove the impulsive noise specific to abovementioned encoding process. Natural images can be represented with less than 0.5 bpp while preserving a reasonable perceptual quality. While both the encoding and the decoding processes require no arithmetic circuits their mixed-signal implementation is extremely simple thus making the proposed scheme very attractive for low power, sensor integrated applications