Performance improvement of edge detection based on edge likelihood index

Abstract

One of the problems of conventional edge detectors is the difficulty in distinguishing noise and true edges correctly using a simple measurement, such as gradient, local energy, or phase congruency. This paper proposes a performance improvement algorithm for edge detection based on a composite measurement called Edge Likelihood Index (ELI). In principle, given a raw edge map obtained from any edge detectors, edge contours can be extracted where gradient, continuity and smoothness of each contour are measured. The ELI of an edge contour is defined as directly proportional to its gradient and length, and inversely proportional to its smoothness, which offers a more flexible representation of true edges, such as those with low gradient, but continuous and smooth. The proposed method was tested on the South Florida data sets, using the Canny edge operator for edge detection, and evaluated using the Receiver Operator Characteristic curves. It can be shown that the proposed method reduces Bayes risk of ROC curves by over 10% in the aggregate test results.