Error measurement for segmentation techniques

In early February of 1994, I decided that I had passed all my exams and that it was time to end my study Computer Science at the Katholieke Universiteit Nijmegen and get my masters degree. I w alked into Theo Schoutens room, to ask for an assignment f o r m y master thesis. He said that he was doing research on setting optimal thresholds for segmentation techniques and he thought imitating satellite images in order to be able to do performance measures on segmentations of those images would be a nice idea. He gave me a PhD thesis and some books on remote sensing and some master thesises on segmenting satellite images, in order to let me know something about the eld of use of my assignment. I started reading, got interested and started to develop and implement t h e method described in this thesis. The goal of my project was to measure error rates in segmentations. To d o this, an artiicial satellite image should be generated given a human made drawing. This segmentation should serve as an input for a segmentation program that was under research. The segmentation program used was based on edge detection and region growing. If diierent thresholds were set to the segmentation program, the region growing part made diierent decisions in whether to merge regions or not. The human made drawing should serve as a correct segmentation. Comparing the correct segmentation to the computer made segmentation, should result in an error measurement of the computer made segmentation. If diierent segmentations were made using the same satellite image, and the error rate of these segmentations were made, then the segmentations could be compared. The segmentation that had the best performance, was made by the segmentation program, when it had the better threshold settings. If the artiicial image now w as similar to real satellite images, then there was a good chance, that the threshold settings that gave t h e best performance in segmenting the artiicial satellite image, would give the best performance in segmenting real satellite images. In the project, I chose to have a exible planning. The programs used in the method should be designed, implemented and tested, and this thesis should be written. I chose to integrate the design of the used programs in this thesis and have a w orking prototype of the programs in a short time. The …