像素点可信度划分层级且具梯度权重的GrabCut  

GrabCut with Gradient Weight and Hierarchical Pixel Credibility

作　　者：沈雅婷 宗平 曾璐洁 黄延浩 SHEN Yating;ZONG Ping;ZENG Lujie;HUANG Yanhao(Nanjing University of Science and Technology Zi Jin College,Nanjing 210023)

机构地区：[1]南京理工大学紫金学院,南京210023

出　　处：《计算机与数字工程》2025年第1期221-227,共7页Computer & Digital Engineering

基　　金：江苏省大学生创新创业项目(编号:202413654018Y);江苏高校“青蓝工程”资助。

摘　　要：对经典前景主体分割算法GrabCut进行研究,其所有的像素点初始化均被标记为前/背景标签,后续均同等参与GMM。但像素点可能处于前/背景边界区域甚至相反类中,将这些像素点的标签传播到它们的近邻像素点,可能会误导分割。因此提出像素点可信度划分层级且具梯度权重的GrabCu(t简称HCGW_GrabCut),初始化通过隶属度聚类(软聚类),如KFCM_II,获得像素点位于非边缘区域的概率,即可信度,再根据可信度对像素点集划分层级并使用梯度权重,使可信度越高的像素点参与后续GMM的成分越高,即参与频率与其权重越高,强调了高可信度像素点的贡献和重要性。最后在背景复杂,前、背景相似度极大或主体为类锯齿状边缘的图像(来源COCO_test2014、DIV2K、BSDS300)上的实证结果显示,HCGW_GrabCut对比GrabCut具有一定竞争力。The article studies the classic foreground subject segmentation algorithm GrabCut,where all pixels are initialized and labeled as front/background labels,and subsequently participate equally in GMM.However,pixels may be located in the front/background boundary region or even in opposite classes,and propagating the labels of these pixels to their neighboring pixels may mislead segmentation.Therefore,the article proposes GrabCut(HCGW_GrabCut),which divides pixel credibility into hierarchical levels and has gradient weights.Initialization is achieved through membership clustering(soft clustering),such as KFCM_iI,to ob⁃tain the probability of pixels being located in non edge regions,which is called credibility.Then,the pixel set is divided into levels based on credibility and gradient weights are used,so that pixels with higher credibility participate in the subsequent GMM compo⁃nents more frequently and with higher weights,emphasizing the contribution and importance of high credibility pixels.Finally,em⁃pirical results on images with complex backgrounds,high similarity between the foreground and background,or subjects with jag⁃ged edges(sourced from COCO_test2014,DIV2K,BSDS300)show that HCGW_GrabCut has a certain competitiveness compared to GrabCut.

关 键 词：图像前景主体分割 可信度 层级 权重 GRABCUT 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]