19/01/2026 3:14 PM - PACES
Aidan Dempster - adempst - 6125 9596
Problem
Identifying the pixels in an image which correspond to what humans intutively call an edge.
Approach
The author notes that formulating the inutive objectives that define an edge mathematically allow for optimization approaches that provide optimal detectors for edges.
They propose that there are three criteria that define a pixel as an edge:
1. Low error rate in both missed detections and spurious detections.
2. Low distance betwen the detected edge and the center of the true edge.
3. A newly proposed criterion, that a single edge should not produce multiple responses.
They define mathematical expressions that capture the first two criteria and then find optimization procedures that can produce filters to localize edges. They subsequently define a denoising procedure that minimizes the prevalence of multiple responses to a single edge. They find that gaussian and derivate of gaussian features well approximate optimal filters.
Contribution
The use of optimization to directly minimize a cost to produce filters that operate effectively given a combination of noise and edge type allow the author to generalize the concept of edge detection that previous research had defined and construct optimal filters for generic edge types. They solved an existing problem in a new way.
Evaluation
They use visuals demonstrating the performance on specific images, a single table showing a small number of performance metrics, and proofs of the optimality of their approach given their stated assumptions.
Substantiation
Based on the evaluation, I understand the theoretical benefits over prior methods that did not use optimization approaches.
I can visually see that the method works.
Cannot see same evaluation applied to other methods. Cannot see how bad prior methods are at the single response criteria proposed by this paper.
A_Computational_Approach_to_Edge_Detection.pdf
End time: 19/01/2026 4:13 PM
Did it in an hour. Good.