Compact image compression using simplicial and ART neural systems with mixed signal implementations

This paper introduces a highly efficient yet easy to implement mixed-signal hybrid neural system for still image compression. The novelty of our proposed structure consists in combining a compact and fast fuzzy-ART (fuzzy adaptive resonance theory) vector quantifier with an adaptive neural system tuned to remove the quantization error. Simulations of our system show that acceptable quality images can be transmitted at rates of about 0.4 bits per pixel (bpp) with a very compact hardware implementation (about 10/sup 3/ devices), i.e. almost 100 times less than actual solutions based on digital implementations of DCT or wavelet transforms. For the sub-QCIF image format (e.g. 64/spl times/64) these rates lead to the possibility of sending video streams (of less than 30 frames per second) at rates below 64 Kbit per second using a cheap technology.