Improving the Canny Edge Detector Using Automatic Programming: Improving the Filter

We have used automatic programming, a machine learning technique related to inductive logic programming and genetic programming, to make the Canny edge detector better at identifying contours in natural images. We present an improved version of the filter used in the first stage of the Canny algorithm. We show that the mean performance of the Canny algorithm with the improved filter on a popular test set of natural images has been improved by 1.4%. Our result shows that the heuristic design provides a statistically significant increase in performance - without adding extra processing steps or adding additional information. This suggests that the filter should be used as a standard part of image analysis platforms. The inferred heuristic filter exhibits an ability to retain detail without sacrificing noise reduction. This is further evidence that automatic programming is well suited for generating heuristics for image analysis problems.