Historical Experiment and Results

Using the new method with a pre-determined, fixed Gaussian filter, the edges shown in the result images were detected automatically after specifying the value of one input parameter to the program. The value of this parameters can be considered as a threshold or a reflection of the researcher's definition of the edge he is trying to detect. If a high valued is specified, it means that the researcher is trying to detect the most prominent edges in the image. But, depending on the goal of the experiment, edges less obvious also be useful and, in this case, a low value should be specified as the input variable. Obviously, the former case is always easier since the S/N ratio is higher. There is always a trade-off between the correctness and completeness of the edges detected. Here, again, the "correctness" and "completeness" are subjective concepts of each researcher.

Several authors at University of South Florida compared the results from edge detection methods proposed by Canny, Nalwa, Iverson, Bergholm and Rothwell. Images from four categories were used and the results are available at their web site. For comparison, one image from each of their four categories was used for testing the new method, as shown in the following examples.

From the experiment results, it can be seen that this improved method has limited spatial resolution because of the finite size of an LMC (the same as any other edge detector). Two side-by-side edges of a linear feature that are very close (less than three pixels) to each other can not be both detected satisfactorily. In some cases, one is detected. In other cases, both are lost. When this happens, it is better to extract the linear feature itself. For example, the linear features whose edges may not have been detected in Example I can be detected directly.

In Example III, at the tip of one of the leaves in the upper left part of the result images, there is a highlighted pixel segment that corresponds to a not-very-obvious edge in the original image. If a threshold is set for the contrast of the edges detected, results can be obtained that does not include such "run-off" edges.

Similar results from these images can be obtained by using EdgeDetector

Edge Detection Home Page
Results from the head image || Results from the Lena image
Download EdgeDetector || Download Synthetic Test Images

Other Projects