Comparative Studies of Region-Based Segmentation Algorithms on Natural and Remote Sensing Images

Abstract

Region-based segmentation algorithms are used as a preprocessing approach to generate over-segmented regions. Over-segmented regions refer to the creation of small regions in an image that represent no meaningful object regions. It has been observed that there are limited works on the performance comparison of the region-based segmentation algorithms on both natural and remote sensing (RS) images. Hence, the objective is to compare the performance of region-based segmentation algorithms on natural and RS images with different complexity of object regions of interest (ROIs). There are four algorithms (Felzenszwalb and Huttenlocher (FH), Quick Shift (QS), Compact Watershed (CW), and Simple Linear Iterative Clustering (SLIC)) being compared using two public datasets. The adapted rand error (ARE) and variation of information (VOI) are used for the segmentation evaluations. Generally, the experiments showed that the SLIC achieved better results as compared to the other algorithms for both images with different complexities of ROIs. This is mainly because the over-segmented regions produced by the SLIC adhered to the image object boundaries well than the over-segmented regions generated by other algorithms. However, CW achieved better average ARE than SLIC for RS images because CW has compactness and marker parameters which influence it to achieve better results.