Studies in Image Segmentation Algorithms Based on Histogram Clustering and Relaxation

摘要

The research in this thesis has focussed upon the algorithms and structures that are sufficient to generate an accurate description of the information contained in a relatively complex class of digitized images. This aspect of machine vision is often referred to as 'low-level' vision or segmentation, and usually includes those processes which function close to the sensory data. The bulk of this thesis devotes itself to the exploration of some of the problems typically encountered in segmentation. In addition, a new and robust algorithm is presented that avoids most of these problems. The analysis is carried out through the use of a series of computer-generated tests images with known characteristics. Segmentation algorithms of varying degrees of complexity are applied to each image and their performance is carefully evaluated. It will be shown that even the most sophisticated algorithms that are currently in use often perform poorly when confronted with certain apparently simple images. In particular, it is shown that techniques which rely on histogram clustering often generate gross segmentation errors due to overlap in the distributions of the individual objects in a scene. Moreover, the relaxation processes used to correct these errors are themselves prone to errors, but of a different kind. Both techniques, clustering and relaxation, fail because they are based on information which is too global to be effective in complex scenes.