Superpixels Segmentation via Growing Minimum Spanning Trees and Reassigning Boundary Pixels

Along with the development of high resolution observation system, the spatial resolution of remote sensing images increases greatly. As it comes to the effective intelligent interpretation of high spatial resolution remote data, the traditional way of per-pixel processing is hardly applicable. On the other hand, the Object-Oriented method has shown its necessity and advantage, that is, grouping contiguous image pixels which are in homogeneous regions into segments (what we call superpixels). In this paper, we propose a superpixel algorithm to segment an image by means of generating minimum spanning trees and iteratively reassigning boundary pixels to improve segment accuracy. The proposed approach is compared with a state-of-the-art superpixel algorithm, SLIC [1], in terms of standard metrics, which are boundary recall (BR), under-segmentation error (UE) and achievable segmentation accuracy (ASA).