摘要
对于视觉场景的理解是机器人在未知环境中进行有目的的行动的一项重要能力,图像语义分割能够有效地帮助机器人理解周围场景的语义特性。条件随机场(Conditional Random Field,CRF)是解决语义分割问题的一个重要框架。针对传统条件随机场相邻节点数过于稀疏的局限性,研究了具有稀疏高阶势CRF的图像语义分割算法,提出一种高阶CRF的二次规划(quadratic programming,QP)松弛推理算法。首先,使用来自TextonBoost的一元势、高斯二元势以及由Pn-Potts模型推导得到的高阶势建立能量函数,然后利用高效的QP松弛推理算法来解决高阶CRF的能量最小化问题,最后采用Pascal VOC2012公开数据集进行实验,验证算法的可行性与有效性。实验证明,该算法有效地克服了传统条件随机场局限性,获得了更好的语义分割结果。
The understanding of visual scenes is an important ability of robots to perform purposeful actions in an unknown environment.Image semantic segmentation can effectively help robots understand the semantic characteristics of surrounding scenes.Conditional Random Field(CRF)is an important framework for solving semantic segmentation problems.Aiming at the limitation of the number of adjacent nodes in the traditional CRF,the image semantic segmentation algorithm with sparse high-order potential CRF is studied,a quadratic programming(QP)relaxation inference algorithm for high-order CRF is proposed.First,the energy function is established using the unary potentials from the TextonBoost,the Gaussian pairwise potentials,and the high-order potentials derived from the Pn-Potts model,and then the high-efficiency QP relaxation inference algorithm is used to solve the high The energy minimization problem of the high-order CRF.Finally,Pascal VOC2012 public data set is used for experiments to verify the feasibility and effectiveness of the algorithm.Experiments show that the algorithm effectively overcomes the limitations of traditional conditions and achieves better semantic segmentation results.
作者
高宇田
杨阳
赵广帅
刘智
GAO Yu-tian;YANG Yang;ZHAO Guang-shuai;LIU Zhi(School of Electronics and Information Engineering,Changchun University of Science and Technology,Changchun 130022)
出处
《长春理工大学学报（自然科学版）》
2021年第2期60-66,共7页
Journal of Changchun University of Science and Technology(Natural Science Edition)
基金
吉林省科技发展计划项目(20180201090GX)
吉林省教育厅“十三五”科学技术项目(JJKH20170618KJ)。
关键词
语义分割
条件随机场
稀疏高阶势
QP松弛推理算法
semantic segmentation
conditional random field
sparse higher order Potentials
QP relaxation inference algorithm
