Unsupervised Color Image Segmentation: with Application to Skin Tumor Borders
This document has been relocated to http://scholarsmine.mst.edu/ele_comeng_facwork/1317
There were 17 downloads as of 27 Jun 2016.
The images used in this research were digitized from 35mm color photographic slides obtained from a private dermatology practice and from New York University. The authors compared 6 color segmentation methods and their effectiveness as part of an overall border-finding algorithm. The PCT/median cut and adaptive thresholding algorithms provided the lowest average error and show the most promise for further individual algorithm development. Combining the different methods resulted in further improvement in the number of correctly identified tumor borders, and by incorporating additional heuristics in merging the segmented object information, one could potentially further increase the success rate. The algorithm is broad-based and suggests several areas for further research. One possible area of exploration is to incorporate an intelligent decision making process as to the number of colors that should be used for segmentation in the PCT/median cut and adaptive thresholding algorithms. For comparison purposes, the number of colors was kept constant at three in the authors'' application. Other areas that can be explored are noise removal and object classification to determine the correct tumor object