一种基于双层框架的仿射类图像抠像方法  

A Hierarchical Framework on Affinity Based Image Matting

作　　者：姚桂林[1,2] 赵志杰 苏晓东[1,2] 辛海涛 胡文[1,2] 秦相林 YAO Gui-Lin;ZHAO Zhi-Jie;SU Xiao-Dong;XIN Hai-Tao;HU Wen;QIN Xiang-Lin(School of Computer and Information Engineering,Harbin University of Commerce,Harbin 150028;Heilongjiang Provincial Key Laboratory of Electronic Commerce and Infor-mation Processing,Harbin 150028)

机构地区：[1]哈尔滨商业大学计算机与信息工程学院,哈尔滨150028 [2]黑龙江省电子商务与信息处理重点实验室,哈尔滨150028

出　　处：《自动化学报》2021年第1期209-223,共15页Acta Automatica Sinica

基　　金：黑龙江省自然科学基金(F2018021,LH2019F044);哈尔滨商业大学校级科研项目(18XN021,2016TD001);哈尔滨商业大学青年创新人才支持计划(2019CX02);黑龙江省哲学社会科学研究规划项目(18GLB029)资助。

摘　　要：仿射类抠像方法主要分为KNN(K-nearest neighbor)类和Matting Laplacian类方法,本文结合这2种方法的优点提出了一种基于仿射类的双层次抠像方法.其中,第一层为绝对像素的划分层次或预处理层次,采用了基于KNN类简单权重与相对远距离的搜索方法,并结合初始Trimap未知区域大小无关的方式;第二层为混合像素的计算层次或最终抠像层次,充分利用了第一层计算获得的剩余混合像素的宽度,自适应地调整Matting Laplacian中的颜色线性模型所构成颜色近邻的核宽度.每个层次均按图像的全局颜色重叠程度相应调整合理的搜索范围.本文的实验具备以下特点:1)预处理层次之后采用了若干典型的后续抠像方法,以展现本文方法相比于其他预处理方法对后续抠像操作步骤的优越性和兼容性;2)最终抠像层次引入了若干其他抠像方法,以验证本文抠像方法的优越性.实验表明,相比于其他单层次的仿射类方法,无论对于计算绝对像素还是混合像素,本文方法都可以大幅提升计算结果的准确率.Affinity based image matting methods can be categorized into KNN(K-nearest neighbor)based matting and matting Laplacian based matting,and this paper raises a hierarchical framework on affinity based matting according to the analyses of the advantages of these two popular affinity based image matting methods.The first opaque pixel classification layer,also named as pre-processing layer,employs a relatively far searching fashion based on simple weights in KNN and is spatial irrelevant to the unknown region of the initial Trimap.The second mixed pixel computation layer,also named as final matting layer,adaptively adjusts the kernel size of the color line model in matting Laplacian according to the remaining size of the unknown region.Each layer adjusts proper searching range adaptively according to the overlapping degree between global foreground and background colors.The following distinctions are provided in the experiments.First,several representative matting algorithms are processed after the first layer to show the superiority and compatibility of our pre-processing method over other pre-processing methods.Second,several alternative matting methods are also processed after the first layer to show the superiority of our final matting method over other matting methods.Experimental results show that our approach can greatly raise the solving precisions for both opaque and mixed pixels.

关 键 词：图像抠像 仿射类抠像 Matting Laplacian KNN搜索 颜色线性模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]