An edge detector derived from three criteria
In November 1986 John Canny wrote edge detection down as the optimisation of three quantities: signal-to-noise ratio, accuracy of the located response, and having only one response to a single edge. Solving it numerically, he showed there is an uncertainty principle between detection and localisation, and that the optimal operator is well approximated by the first derivative of a Gaussian.
Why it matters
An edge detector stopped being a matter of taste. The criteria were named, operators became comparable by a number, and two devices from this paper — non-maximum suppression along the gradient and thresholding with hysteresis — still sit in the libraries under the author's name.
The figures the paper prints: the first derivative of a Gaussian is worse than the optimal operator by about 20 percent on the product of criteria, and worse by about 10 percent on the multiple-response measure. The author writes that a difference of this size would be hard to see on real images, and that computing the first derivative of a Gaussian in two dimensions costs much less, which is why it is what the experiments use. The ratio of the high to the low threshold in the hysteresis is in the range two or three to one. The record does not claim an "approximation error under 20 percent": in the paper 20 percent is how much worse the approximate operator performs, not how far its shape departs. The manuscript was received on 10 December 1984 and revised on 27 November 1985. At publication the author is given as being with the MIT Artificial Intelligence Laboratory.