Edge detection using neural network arbitration

A human observer is able to recognise and describe most parts of an object by its contour, if this is properly traced and reflects the shape of the object itself. With a machine vision system this recognition task has been approached using a similar technique. This prompted the development of many diverse edge detection algorithms. The work described in this thesis is based on the visual observation that edge maps produced by different algorithms, as the image degrades. Display different properties of the original image. Our proposed objective is to try and improve the edge map through the arbitration between edge maps produced by diverse (in nature, approach and performance) edge detection algorithms. As image processing tools are repetitively applied to similar images we believe the objective can be achieved by a learning process based on sample images. It is shown that such an approach is feasible, using an artificial neural network to perform the arbitration. This is taught from sets extracted from sample images. The arbitration system is implemented upon a parallel processing platform. The performance of the system is presented through examples of diverse types of image. Comparisons with a neural network edge detector (also developed within this thesis) and conventional edge detectors show that the proposed system presents significant advantages.