Agglomerative oversegmentation using dual similarity and entropy rate

摘要

Oversegmentation is a preprocessing step for many computer vision and remote sensing tasks, such as object recognition and image understanding. In this paper, a method called agglomerative oversegmentation (AOS) is proposed. AOS first designs a dual similarity of the connected pixels by considering both the pixel and region levels. The entropy rate of random walks is then employed on the image plane, and the gain of each pair of connected pixels is estimated. Finally, a fast agglomerative algorithm is developed to oversegment the image using an entropy rate gain function. Two challenging datasets are utilized in the experiments which show the promising performance of AOS. In contrast to the state-of-the-art algorithms, the oversegmentation of AOS is more consistent with the geometric structures of the target objects, especially for the objects with complex texture on remote sensing images. (C) 2019 Elsevier Ltd. All rights reserved.