A 32-neurons Slice for High Speed Handwriting Recognition
暂无分享,去创建一个
This article presents PAPRICA-3, a high-speed 32-neurons slice for real-time neural processing of images. The system is a programmable parallel processor array with the instruction set tailored to the emulation of neural networks and to image processing. Dedicated hardware features allow simultaneous image acquisition, image processing and neural network emulation. The 32-neurons slice is cascadable to match the required size. The system is under fabrication and has been simulated at a clock frequency of 100 MHz (typical). Estimated performance is up to 130 MCPS for perceptrons. The system has been tested on the recognition of handwritten check amounts, by integrating neural network algorithms with context analysis techniques.
[1] Simon Haykin,et al. Neural Networks: A Comprehensive Foundation , 1998 .
[2] Sargur N. Srihari,et al. Off-Line Cursive Script Word Recognition , 1989, IEEE Trans. Pattern Anal. Mach. Intell..
[3] Hans-Leo Teulings,et al. Invariant Handwriting Features Useful in Cursive-Script Recognition , 1994 .
[4] Jean Serra,et al. Image Analysis and Mathematical Morphology , 1983 .