面向自然图像透明区域的特征感知抠图  

Feature-aware matting for transparent regions in natural images

作　　者：张岚岚 ZHANG Lanlan(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang 550025,China;Key Laboratory of Pattern Recognition and Intelligent System of Guizhou Province,Guizhou Minzu University,Guiyang 550025,China)

机构地区：[1]贵州民族大学数据科学与信息工程学院,贵阳550025 [2]贵州民族大学贵州省模式识别与智能系统重点实验室,贵阳550025

出　　处：《智能计算机与应用》2024年第12期139-144,共6页Intelligent Computer and Applications

基　　金：贵州省科技计划项目(QKHJCZK2022YB195,QKHJCZK2023YB143,QKHPTRCZCKJ2021007);贵州省教育厅自然科学研究项目(QJJ2023061,QJJ2023012,QJJ2022015)。

摘　　要：自然图像抠图作为计算机视觉的重要任务之一,是一种通过预测精确的alpha matte来将前景物体从图像中分离出来的技术。由于透明区域前背景像素高度混合,在提取透明区域遮罩时易出现误抠问题。为了解决这一问题,本文提出透明特征感知抠图模型,模型中带有一个专门为透明区域遮罩误抠设计的透明特征感知模块。该模块能够多尺度提取上下文信息,精确感知并分离前背景像素高度混合的透明区域,同时提取透明的细致特征,从而有效解决了自然图像中透明区域的遮罩误抠问题。在Composition-1k测试集上的实验结果表明,透明特征感知抠图模型在解决透明区域遮罩误抠问题上具有一定的优越性。与经典的DIM抠图方法相比,本文方法在MSE、SAD、Grad和Conn四个评价指标上分别提升了45.71%、30.95%、47.74%和39.96%。Natural image matting,as a critical task in computer vision,involves separating foreground objects from images by predicting precise alpha mattes.Due to highly blended foreground-background pixels in transparent regions,errors in mask extraction are common.To address this issue,this paper proposes a transparent feature perception matting model equipped with a transparent feature perception module,which is specifically designed to mitigate mis-matting in transparent region masks.This module extracts contextual information across multiple scales,accurately perceives and separates highly blended foreground-background pixels in transparent regions,and extracts detailed transparent features,effectively resolving mis-matting issues in natural images.Experimental results on the Composition-1k dataset demonstrate the superiority of the transparent feature perception matting model while addressing mis-matting issues in transparent region masks.Compared to classical DIM matting methods,the proposed approach shows improvements of 45.71% in MSE,30.95%in SAD,47.74% in Grad,and 39.96% in Conn evaluation metrics.

关 键 词：图像抠图 透明区域 感知 上下文信息 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]