结合DBSCAN聚类的室内场景分割  

Segmentation for Indoor Scenes Based on DBSCAN Clustering

作　　者：刘梦迪 潘晓 高珊珊[2] 辛士庆 周元峰 Liu Mengdi;Pan Xiao;Gao Shanshan;Xin Shiqing;Zhou Yuanfeng(School of Software, Shandong University, Ji’nan 250101;School of Computer Science and Technology, Shandong University of Finance and Economics, Ji’nan 250014;School of Computer Science and Technology, Shandong University, Qingdao 266237)

机构地区：[1]山东大学软件学院,济南250101 [2]山东财经大学计算机科学与技术学院,济南250014 [3]山东大学计算机科学与技术学院,青岛266237

出　　处：《计算机辅助设计与图形学学报》2019年第7期1183-1193,共11页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61772312,61772016);山东省重点研发计划项目(2017GGX10110,2017GGX10109);山东省自然科学基金(ZR2017MF033);山东大学基本科研业务经费资助(2018JC030)

摘　　要：针对 RGB-D 图像具有丰富的三维几何特征,复杂度高这一具有挑战性的难题,提出一种针对室内场景RGB-D 图像的分割算法.首先,经过 RGB-D 图像过分割生成超像素,并基于超像素之间的距离度量测量超像素之间的相似性;然后,采用 DBSCAN 算法将具有相似的颜色信息和几何信息的超像素聚类到一个分类中.在该聚类过程中,通过限制扩散区域来降低计算复杂度.在室内场景 RGB-D 图像库上大量实验结果表明,文中算法分割精确度和速率均超过了其他算法,证明了其高效性和准确性.Aiming at the challenging problems of RGB-D images with rich 3D geometric features and high com- plexity, this paper proposes a segmentation algorithm for RGB-D images of indoor scenes. Firstly, generating superpixels by over-segmentation of RGB-D images and measuring the similarity of two superpixels based on the distance. Then, the DBSCAN algorithm is used to cluster the superpixels with similar color and geo- metric information into the same classification. In the clustering process, we restrict the diffusion area to reduce computational complexity. A lot of experimental results on the database of RGB-D images show that the segmentation accuracy and rate of our algorithm exceed the other algorithms, which proves our algo- rithm’s efficiency and accuracy.

关 键 词：RGB-D 图像 超像素聚类 DBSCAN 图像分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]