Supervised region merging based on binary classification and active learning strategy for segmentation of very high resolution imagery

Abstract

Image segmentation is a prerequisite and important step of object based image analysis. The quality of image segmentation directly affects the subsequent object based classification and analysis. Region merging is an important step of many image segmentation methods. Most existing region merging methods use spectral data and a user-defined threshold. The determination of the threshold for region merging is usually done by a trial and error process. In this study, we proposed a new region merging method. The method views the task of deciding if two adjacent segments should be merged as a binary classification problem. The support vector machine (SVM) is adopted as the binary classifier used in this study. To collect appropriate training samples used in SVM classification, active learning strategy was adopted to obtain more informative training samples. The proposed method was validated by comparing with existing threshold based region merging method. The experimental results demonstrate that the proposed method outperformed the existing region merging method.