Color segmentation of aerial images by fuzzy clustering

Abstract

The method of fuzzy clustering was applied to color images, where the color in RGB space is the attribute used for segmentation. Fuzzy c-means was employed, with the Mahalanobis distance metric to detect elongate groups of pixels with similar colors. The optimal number of classes was found by minimizing the Xie-Beni index. From the main image, sub-images were reconstructed with pixels classified in the same group. As a result, a good separation was obtained between visually different environmental structures such as forests, water, soil and other. As color is invariant under image rotation and magnification, this method is very robust to perform separation between the different regions. One can carry out post-processing to merge semantically similar sub-images.