The regularization-based approach discussed in the previous section is not the only way of using local edge context to find reliable image edges. Two other methods of using this information are described in [80] and [47].
In [80] Otte and Nagel find the edge gradient response in a Gaussian smoothed image. Then, after non-maximum suppression, a rule-based scheme is used to validate potential edges; the local gradient direction is analyzed, and edge-like continuity in orientation is required for the edge to be reported. This context-based edge thresholding process is used instead of the more usual thresholding based on gradient magnitude. Results suggest that this method can give cleaner output than simple one level thresholding, but connectivity at junctions is not improved.
In [47] Haralick and Lee find the probability of a pixel being an edge point by analyzing possible edge routes through the pixel's locality. An evaluation function for the edge quality of each possible route is defined, and the optimal route is calculated. The method is tested on very simple synthetic images; it is not apparent how well this method would work on complicated (i.e. real) images.