This thesis describes an image segmentation technique based on watersheds, a clustering technique which does not use spatial information, but relies on multispectral images. These are captured using a monochrome camera and narrow-band filters; we call this color segmentation, although it does not use color in a physiological sense. A major part of the work is testing the method developed using different color images.
Starting with a general discussion of image processing, the different techniques used in image segmentation are reviewed, and the application of mathematical morphology to image processing is discussed. The use of watersheds as a clustering technique in two- dimensional color space is discussed, and system performance illustrated. The method can be improved for industrial applications by using normalized color to eliminate the problem of shadows. These methods are extended to segment the image into regions recursively. Different types of color images including both man made color images, and natural color images have been used to illustrate performance. There is a brief discussion and a simple illustration showing how segmentation can be used in image compression, and of the application of pyramidal data structures in clustering for coarse segmentation.
The thesis concludes with an investigation of the methods which can be used to improve these segmentation results. This includes edge extraction, texture extraction, and recursive merging.