摘要
针对目前在对图像中的物体进行分割时存在的分割精度不高,分割后物体缺失严重,边缘不清晰等问题,提出了一种结合多种图像分割算法的实例分割方案.该方案首先通过具有实例分割功能的Mask RCNN算法对输入的图像进行初步的分割,得到初始掩膜.再通过SLIC超像素分割算法对原图进行超像素分割得到超像素块,结合超像素块对初始掩膜的边缘进行扩展,结合扩展后的掩膜和初始掩膜进行形态学操作得到GrabCut算法分割的初始三元图,该图中明确指出了确定的前景、确定的背景和待分割区域,在此基础上用改进的GrabCut算法建立高斯混合模型(Gaussian Mixed Model,GMM),并反复迭代高斯混合模型参数,最后利用最大流最小割算法得到最优目标分割结果.实验结果表明,本文所提出的图像分割方案,分割效果在直观上能保证物体的完整性,基本能够将物体的所有信息都分割出来.与其他的分割算法进行比较,本文方案在分割精度上平均提高了约6.9%,同时具有很好的视觉效果.
Based on multiple image-process methods,an instance segmentation method has been proposed to get masks which are more accurate than typical methods.First,M ask RCNN is employed to get initial masks.Second,by using SLIC super-pixel segmentation method,the initial masks can be enlarged by super-pixel blocks.Third,through morphological operations,a three-item image can be calculated from the initial and enlarged masks,in which pixels have been classified into the foreground,the background or unsure.Forth,an improved GrabCut algorithm uses the three-item image as initial,and iteratively optimize its Gaussian M ixed M odel.At last,the final masks can be computed by the max-flow min-out method.The experiment results show that the proposed method can improve the accuracy by 6.9%.The output masks are sharp and complete.
作者
詹琦梁
陈胜勇
胡海根
李小薪
周乾伟
ZHAN Qi-liang;CHEN Sheng-yong;HU Hai-gen;LI Xiao-xin;ZHOU Qian-wei(College of Computer Science&Technology,Zhejiang University of Technology,Hangzhou 310000,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2020年第4期837-842,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金青年项目(61802347)资助。
