The Community for Technology Leaders
Green Image
We describe an image analysis system developed around two NET32K neural network processors. This single-board system scans simultaneously 64 templates of size 16x16 pixels over bilevel images. For each template it produces a feature map, marking where in the image a match has been found. The system processes over 20 images of size 512x512 pixels in one second. Two analog neural network chips are at the heart of the system, computing over 100 billion multiply-accumulates per second. By coding simple generic shapes, such as strokes, corners, and end-lines into the templates, the resulting feature maps describe the image's content with a set of building blocks. From this representation a quick and robust analysis of the image's content is possible, even if the images are complex and noisy. The system is used for such tasks as locating address blocks on mail pieces and analyzing the layout of checks.
Neural networks, image analysis, neural nework processors
Hans Peter Graf, Eric Cosatto, "A Neural Network Accelerator for Image Analysis", IEEE Micro, vol. 15, no. , pp. 32-38, June 1995, doi:10.1109/40.387680
92 ms
(Ver 3.3 (11022016))