Back
to indexJohn Canny
Summary
This paper is the foundation upon which modern edge detection algorithms are generated. "Canny edges" are commonly referred to in image processing programs and math programs like mathematica. Two concrete goals are identified: detection of edges with a low error rate and localization performance. Localization performance means the distance between the points marked by the detector and the "center" of the true edge should be minimized. The author notes there is a natural uncertainty principle between the two goals: the more refined the localization, the more noise is produced, increasing the error rate. The inverse is also true, where the error rate may be reduced by reducing the data that is fed into the localization routines. I must admit the mathematics behind the technique are too intricate for me to explore under limited time constraints.
Methods
First, we define the signal to noise ratio of the image, which we would like to maximize. Then, filters are applied that respond to different noise patterns in the image. The filters that are effective at discriminating between the signal and the noise at the center of the edge are identified (by a complex formula (9)). Optimizations are performed based on the sampled operator impulse response, or the distance between adjacent noise maxima. This data is input into a scaling function to smooth out the points, and separate out a single unique solution. The author points out there is no best scaling function, but rather a class of functions that arbitrarily fit the noise patterns. Probability of false edges are reduced, by their Gaussian properties.
Keywords
edge detection, image processing, feature extraction, machine vision, multiscale image analysis
Rating
6
Bibtex Entry
@article{ canny86computational,
author = "John Canny",
title = "A Computational Approach to Edge Detection",
journal = "IEEE Transactions on Pattern Analysis and Machine Intelligence",
volume = "8",
number = "6",
pages = "679--697",
year = "1986",
url = "http://www.cs.berkeley.edu/~jfc/papers/papers.html"
}