We are currently optimizing the edge and line detectors involved in our image processing algorithms. We adjust various parameters involved in the process of edge and line detection and look for the optimal parameter set which helps generate the best performance.

First of all, we created reference data by manually labeling the image database. For this purpose, we modified our front-end image segmentation tool, incorporating the function of drawing lines. (We could only draw polygons previously.) We labeled two sets of images from the Pre-Phase 01 database. The first set consists of 165 images, and the second one contains 159 images. Here is an example of the manual work.

We designed a metric to measure the performance of the detectors. In this scheme, detected lines are evaluated according to how close they are to the corresponding reference lines in position, length and slope. To be specific, for each reference line, we find a detected line which is the closest to it in position. Then we compute the distance from the detected line to this reference line, compare the lengths of both lines, and compare the slopes of them. If all these results are within the prescribed thresholds, we accept the detection as a correct one. Those detected lines without any reference lines to match them are considered as insertion errors.

Some preliminary evaluation results and example images are shown here. We note that in most cases, detected lines were considered incorrect detection because they were much shorter than their references. This is actually a common problem called "broken lines" in line detection. Also, there were many inserted lines, which may be reduced by an adjustment of the threshold parameters. The problems will be investigated further in the coming experiments.