Improving the Canny Edge Detector Using Automatic Programming: Improving Hysteresis Thresholding

Abstract

We have used automatic programming to improve the hysteresis thresholding stage of the popular Canny edge detector—without increasing the computational complexity or adding extra information. The F-measure has been increased by 1.8% on a test set of natural images, and a paired student-t test and a Wilcoxon signed rank test show that the improvement is statistically significant. This is the first time evolutionary computation and automatic programming has been used to improve hysteresis thresholding. The new program has introduced complex recursive patterns that make the algorithm perform better with weak edges and retain more detail. The findings provide further evidence that an automatically designed algorithm can outperform manually created algorithms on low level image analysis problems, and that automatic programming is ideally suited for inferring suitable heuristics for such problems.