基于深度图及支撑分析的图像分割  

Image Segmentation Based on Depth Map and Support Analysis

作　　者：皮志明[1] 汪增福[1,2] 

机构地区：[1]中国科学技术大学自动化系,合肥230027 [2]中国科学院合肥智能机械研究所,合肥230031

出　　处：《小型微型计算机系统》2014年第6期1342-1347,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(60875026)资助

摘　　要：通过分析场景区域之间的支撑关系,提出一种分别在支撑区域与非支撑区域作深度图分割的图像分割算法.现有分割算法往往基于整个场景作不同物体深度不连续性假设,忽略了场景内不同物体之间的支撑关系,使得这类算法无法将支撑物与被支撑物分割开.该文首次运用深度图分析三维场景表面法向量并将法向量近似垂直于水平面的区域判为支撑区域.根据不同物体深度不连续性假设,分别在支撑区域与非支撑区域作深度图分割,并将支撑区域与其下方在深度上连续的非支撑区域合并.最后用彩色图像的过分割边界来修正深度图像分割边界.该算法能够有效地将支撑物与被支撑物分割开,将相互遮挡的不同物体分割开,在三个公共数据集上的测试结果表明了该算法在各种场景下的有效性和鲁棒性.This paper proposes a novel image segmentation algorithm by segmenting depth map in support regions and non-support regions separately. Support regions and non-support regions are acquired by support relation analysis between different regions in a scene. Most existing works assume depth discontinuity between different objects in a whole scene. These works also ignore support relation between different objects. Their assumptions make it hard to separate different objects with support relations. In this work, the surface normal vectors of the scene are calculated using the depth map and regions with vertical normal vector are set to be support regions. Depth map is segmented in support regions and non-support regions separately based on depth discontinuity. Then segmentation results in support and non-support regions are combined together. At the same time, support segments are merged with non-support segments when non-support segments are below the support segments and the boundaries are continuous between them in depth. Finally, Depth segmentation edges are forced to be consistent with the color over-segmentation. Experiments on three public datasets demonstrate our validity and robustness to segment support and non-support regions apart and to separate different objects with occlusion boundaries between them.

关 键 词：图像分割 深度图 支撑系 深度不连续 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]