期刊名称：ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
印刷版ISSN：2194-9042
电子版ISSN：2194-9050
出版年度：2010
卷号：XXXVIII - 4/C7
出版社：Copernicus Publications
摘要：The relation between the segmentation parameters values and the segmentation result is far from being obvious. Therefore, in order to produce the desired outcome a complex and time consuming trial and error process is usually required. Automatic methods based on Genetic Algorithms (GA) have been proposed that endeavor to adjust automatically the segmentation parameters to a given set of reference segments manually delineated by a human analyst. The method searches the parameter space for a set of values that optimizes a given fitness function, which should express numerically the similarity between the segmentation outcome and the reference segments. The fitness functions proposed for that purpose were designed so that they achieve their extreme value when a perfect match with the reference is produced. However, there is no theoretical foundation as well as no experimental study that confirms the adequacy of these adaptation methods when a perfect match is not possible. This corresponds to most practical applications, in which the best attainable outcome differs from the reference, and the obtained similarity value departs from the ideal one. This work addresses these issues and investigates the performance of the GA based adaptation methods for a number of different similarity metrics on different types of reference objects. Working on a Quickbird test image, the study compares the different metrics and examines their correlation degree. The work lastly assesses if these metrics lead the GA to the same solution and ultimately verify the assumptions underlying the GA adaptation method
