This page describes a primary motivation for profile-based edge detection.

"Edge detection is one of the most commonly used operations in image analysis, and there are probably more algorithms in the literature for enhancing and detecting edges than any other subject.''
-- J.R.Parker in "Algorithms for Image Processing and Computer Vision", John Wiley and Sons, 1996, page 1.

One of the reasons we propose another edge detector is that most edge detectors are unsatisfactory. Typically, they produce value for each pixel in the input image. The value indicates the likelihood that the pixel is on an edge. A threshold is then used to decide which pixels are edge points. There are several problems with this approach:

1. The selection of the threshold is arbitrary. Different images require different thresholds. Sometimes, even different areas within the same image would need different thresholds. There are not many methods for automatic selection of the threshold (and these ought to be an integral part of each detector).
2. The results from these detectors are just points, while most higher level image processing operations, especially the model-based ones, require the edge curves.
3. Edges must be characterized by their profiles -- but almost none of the edge detectors attempt to output such profile information. Hueckel (1979) is an exception.

Note the three subjects in the problems we show here can be actually intertwined. The precision of the threshold selection and the profile characterization can affect each other, as the profile characterization requires some noise tolerance which is bounded by a threshold, while the threshold of the quantumized edge magnitude can make use of the noise measurement abtained from profiles. As for edge linking, the hysteresis thresholding of the widely-used Canny detector uses edge linking as part of the thresholding process, even though in implementation, most people use image traversal to replace the edge linking process. On the other hand, almost no edge linking processes employ edge profiles, even though the edge profiles can generally provide better description of the edges, therefore, yield better edge linking results. A few examples are shown below:
Original images:

An edge using Canny detector (the edge and the start point is marked in each image) :

In the first image, it can be said either correct or not depending on the point of view. Some would argue that the edge should not continue at the position of the junction. More importantly, there is no control on whether the edge should continue or not. And for the second edge, the linking is definitely wrong. The edge of the floor tile should be separated from the edge of the box. In comparison, the following two images are the results from our experimental profile-based edge detectors:

In these results, the edges are stopped where the profiles begin to change significantly.

Our new proposed method will address the above issues. Here is further information on our research: