Clustering is usually formulated as an optimization problem with objective functions which have several local minima. Real image segmentation techniques using numerous control parameters may not apply several images uniformly because the parameters interact in a nonlinear fashion. Therefore, most of results are not always crisp nor correct
In this thesis, we propose new clustering and image segmentation method based on genetic algorithm to solve those problems. The genetic algorithm is used as a tool to search a good or usable clustering and image segmentation, which maximizes the quality of regions or clusters generated by split-and-merge processing.
For clustering, we propose new measure function based on structural relationship by the nearest neighbor clusters, and by a degree of relative separations among clusters. Furthermore, we propose an objective function for image segmentation which measures a degree of separation and compactness between and within finely segmented regions, and an edge strength along boundaries of all regions. These measures based on the fuzzy decision by fuzzy membership function, might be used in many applications such as pattern recognition, classification and image understanding as well as to image segmentation.
To efficiently apply the clustering and image segmentation, we newly modify several operations of the existed genetic operators, and propose new genetic model which is elite-based subpopulation model to improve the performance of the genetic algorithm.
We present several experimental results to demonstrate the capability of the proposed approach. The new approach provides useful results without the need for critical parameters or threshold values, iterative visual interaction, or a priori knowledge of test pattern and image.