|
|
|
“An Efficient Digital VLSI Neural Processing Element Design for Image Processing”
by Josephine Chia-Fen Chang
October 1994
With increasing computing power of a chip implemented by silicon technology, an intelligent machine which possesses basic skills of sensing, signal processing, and moving based on human brain models becomes feasible in the near future. In addition, artificial neural networks have the potential to solve many complex and time-consuming engineering and scientific problems with inherently massively parallel processing architectures. To obtain an optimized solution from a neural network, the paralleled hardware annealing method can be applied. The results on neural networks with multi-level nonlinearities are presented. Hardware using parallel architecture could greatly speedup neural network operations. By using the digital processor design approach, high-precision requirements for neural network algorithms can be easily achieved. A custom-designed digital VLSI processing element (PE) for general-purpose neurocomputing is presented. Detailed communication networks, instruction sets and circuit blocks are created for the one-dimensional ring-connected and two-dimensional mesh-connected systolic array. The reduced instruction set technique and microprogramming skills can be applied to optimize the software control of the processor array. A prototype PE has been designed and fabricated in a 6.19 x 5.46 mm2 microchip by using the 0.8-µm CMOS technology from Hewlett-Packard Company through the MOSIS Service. By arranging the PE layout in a ring-connected array architecture, a 20-PE chip is estimated to occupy a silicon area of 2.09 x 1.93 cm2 by using a 0.5-µm CMOS technology. A digital signal processor chip can be used to broadcast or pipeline microcodes to all processing elements in an array. This neural PE design is suitable for image processing. System-level simulation results of applications including printed character recognition and image compression based on neural network algorithms using the designed hardware are also presented.