A trainable analog neural chip for image compression

Video motion estimation and high-ratio image compression are two key data processing steps for advanced television systems and imaging machines. The detailed circuit design of a vector quantization chip with the full-search scheme is presented. Each 5*5 image block with 256 pixel-gray levels can be quantized into the 6-bit codebook by a prototype chip of 4.6 mm*6.8 mm in a 2- mu m scalable MOSIS technology. A speedup factor of 750 over a Sun-3/60 workstation has been obtained. Adaptive codebook learning can be performed with a digital coprocessor.<
>