CAJ | 학술논문

针对现有的交互式分割算法对用户标记的数量与位置敏感的问题，提出了一种与图像区域特性结合的分割算法．首先将均值漂移分割的小块区域作为虚拟像素，参与有权图构建，从而为信息向远方同质区域的像素传播提供路径；然后设计一个基于像素归属于各类用户标记的概率代价函数，并采用微粒群算法进行全局寻优，最后将获得的归属概率作为判据实现分割．实验结果表明，即使采用少量用户标记也可以获得满意的分割结果，说明该算法对标记的数量与位置具有很高的鲁棒性．
침대현유적교호식분할산법대용호표기적수량여위치민감적문제，제출료일충여도상구역특성결합적분할산법．수선장균치표이분할적소괴구역작위허의상소，삼여유권도구건，종이위신식향원방동질구역적상소전파제공로경；연후설계일개기우상소귀속우각류용호표기적개솔대개함수，병채용미립군산법진행전국심우，최후장획득적귀속개솔작위판거실현분할．실험결과표명，즉사채용소량용호표기야가이획득만의적분할결과，설명해산법대표기적수량여위치구유흔고적로봉성．
Existing interactive segmentation algorithms are sensitive to quantity as well as placement of user＇s scribbles. Hence a novel algorithm which can employ region cue is proposed. Firstly, segmentation result by mean-shift algorithm is regarded as imaginary pixels, and further incorporated into construction of weighted graph , which provides long range connection between homogenous semantic areas. Then cost function based on likelihoods to each label is put forward and particle swarm algorithm is utilized to find the solution to this global optimization. Finally segmentation is completed based on resulting likelihoods. Experiment results indicate that even little amount of scribble can result in satisfactory visual effect, which demonstrats the proposed method be of high robustness to quantity and placement of scribbles.