The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

MSc Thesis #9520

Title:A Neural Network Vision System Based on the Curvature Primal Sketch
Date: 1995
Abstract:In computer vision, simulated neural networks can be used to classify images. The ;usual method of doing this involves mapping the pixel values of the image onto the input nodes of a feed-forward net. This is problematic in that the topological properties of the original image, such as the spatial relations between different pixels, are not reflected in the net since all input nodes are initially equivalent. This thesis addresses the problem of recognising curves. Extensive preprocessing of the curves identifies the interesting features, following the method of the Curvature Primal Sketch ([Brady and Asada 1989]). Coordinates of these features are used as input values to the net. The net can then use the differences in coordinate values to gain information about spatial relations between arbitrary image features. Also, much of the irrelevant information is removed in the process of identifying the important features, making the recognition task easier. The nets obtained are small and perform recognition well. A net with dimensions 14-14-10 was trained with collection of 500 handwritten digits, collected from different people. On a similar test set of 100 digits, the net achieved 92.8accuracy in recognition.

[Search These Pages] [DAI Home Page] [Comment]