OEA Based Clustering Algorithms And Its Application On SAR Image Segmentations

SYNTHETIC aperture radar(SAR) images find increasingly wide applications because SAR sensors can penetrate clouds and work in bad weather conditions and in nighttime when optical sensors are inoperable. An important problem in SAR image applications is correct segmentation. It is the basis of the understanding of SAR images, such as the change detection of regions for maps updating, the recognition of targets, and so on.Clustering is a baisic technology that can discover the realtionships between data. Based on the defferent degrees of similarity between data, clustering technology will devide data into several groups, and data in the same group have great similarity while data from different groups have little similarity. And exploiting the ability of clustering to segmentate SAR images has now been a hot research area. Following is our works in this domain:1. The FCM clustering algorithm is easy to be traped in local optimal and is sensitive to the initialization. Besides, no special information is utilized when FCM is applied in image segmentation. To solve these problems, we proposed a new hybrid algorithm called OEA-FCM that utilizes the global searching ability of the Organizational Evolutionary Algorithm(OEA) to solve the drawback of FCM. Moreover, spatial information is introduced in OEA-FCM to improve the robustness to noise.2. The Iterative Self-Organizing Data Analysis Technique Algorithm(ISODATA) is a typical clustering algorithm for image segme ntation which can automatically adjust the number of clusters and cluster centers via the division and merging of clusters. However, the parameters for division and merging are crucial and the performance of the algorithm is sensitive to the selection of these parameters. To deal with this problem, we proposed a new hybrid algorithm(OEA-ISODATA) that combines the OEA and ISODATA, in which the global searching ability of OEA is utilized to search for the optimal parameters for ISODATA. Moreover, clustering is conducted on image patches instead of pixels in order to introduce the spatial information and hence improve the robustness to noise.3. Since clustering validity index only takes certain properties of clustering result into consideration, and no one clustering validty index could fit all kinds of data. Therefore the ability to search optimal solution is limited when it comes to the single objective EAs that sets only one validity index as the fitness function. To solve this drawback, a clustering algorithm based on the direction-based multiobjective evolutionary algorithm(DMEA) is proposed, and it is named DMECA. Both the diversity and convergency is considered when genretate new individuals in DMECA, and a archive is constructed to maintain the nondominated solutions. Besides, strategies are designed to make nondominated solutions uniformly distributed at pareto front when update the archive and the population.