A Combined Statistical/Neural Network Multi-Scale Edge Detector

I. Williams, N. Bowring, E. Guest, P. Twigg, Y. Fan, and D. Gadsby (UK)


Edge detection, statistical, multi-scale, textured and noisy images.


A method of performing edge detection on noisy and textured images is presented. The edge detector utilises two-sample statistical tests such as the Student's T-test or F-Test test to compare two regions surrounding a particular pixel. By applying a series of two-region masks at varying angles, the pixel can be classified as edge or non-edge with the likely edge direction also being determined. The output of several of these two-region masks of differing sizes is applied to a neural network classifier, trained to detect edges only present at a range of scales. The performance of the edge detector is compared to commonly used or comparable algorithms such as the Canny and Susan edge detection filters. The results show that this filter compares well and robustly produces edge information on a series of images. The filter has been used to on histological images and x-ray images of passenger luggage and has been found to perform extremely well in highlighting partially concealed objects.

Important Links:

Go Back