基于加权最小二乘的主结构快速提取算法  

A fast main structure extraction algorithm based on weighted least squares

作　　者：刘堂友[1] 于符婷 张笑源 Liu Tangyou;Yu Futing;Zhang Xiaoyuan(College of Information Science and Technology,Donghua University,Shanghai,201620,China)

机构地区：[1]东华大学信息科学与技术学院,上海201620

出　　处：《南京大学学报（自然科学版）》2022年第3期430-439,共10页Journal of Nanjing University（Natural Science）

摘　　要：从复杂纹理图像中提取主结构是计算机视觉和图形应用的基本过程.针对加权最小二乘法依赖于梯度大小、无法去除对图像语义贡献很小的小规模、高对比度的振荡细节(如纹理)的问题,提出一种新的用于抑制图像纹理的权重算子,并对该权重算子的有效性进行验证.为了解决在优化全局目标函数过程中需要求解大型稀疏拉普拉斯矩阵、计算成本高的问题,采用代数多重网格算法作为共轭梯度法的预处理算子加快稀疏矩阵方程的求解速度.实验表明,提出的权重算子能有效地抑制图像纹理,并且图像主结构的边缘不会被模糊,其滤除纹理、提取主结构的效果优于其他同类算法.另外,所用的加速算法和其他传统预处理算法相比,能将主结构的提取时间缩短很多.Extracting the main structure from a complex texture image is the basic process of computer vision and graphics applications. Aiming at the problem that the weighted least squares method(WLS) depends on the size of the gradient and cannot remove the small-scale and high-contrast oscillating details(such as texture) that contribute little to the semantics of the image,this paper proposes a new method for suppressing image texture. In order to solve the problem of large sparse Laplacian matrix and high calculation cost in the process of optimizing the global objective function,this paper adopts the algebraic multigrid algorithm used as the preprocessing operator of the conjugate gradient method to speed up the solution of sparse matrix equations. Experiments show that the proposed weighting operator can effectively suppress the image texture and prevent the edges of the main structure of the image from being blurred. The effect of filtering the texture and extracting the main structure is better than other similar algorithms. In addition,the acceleration algorithm used in this paper can shorten the extraction time of the main structure compared with other traditional preprocessing algorithms.

关 键 词：纹理 主结构提取 加权最小二乘 稀疏矩阵 代数多重网格 共轭梯度 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]