摘要
针对彩色图像分割问题,研究Markov随机场(Markov random fields,MRF)模型内迭代条件模式(Iterative conditional mode,ICM)方法的标记推理策略.通过小波分解构造图像多尺度表达,针对顶层图像先验标记获取问题,改进原始谱聚类算法,通过近邻传播自动确定图像的聚类参数,运用集成学习提高算法的稳定性和准确度.对其他各尺度图像,通过分析尺度关联下的区域特征变化,结合不同尺度间的特征相似性和同一尺度内空间邻域的一致性,提出一种立体结构描述下的尺度–空间映射法则.通过定量和定性的分割实验,结果表明本文算法具有良好的准确性、鲁棒性和普适性.
For the purpose of color images segmentation, the problem of inferring strategy for iterative conditional mode (ICM) algorithm in the Markov random fields (MRFs) is revisited. Using wavelet decomposition to construct multi- solution expressing of image, the original spectral clustering algorithm is improved to solve the problem of acquiring the prior labels for top-scale image. The parameters of image clustering are generated by affine propagation automatically and the prior labels are acquired after optimizing with ensemble learning to improve the accuracy and stability of algorithm. By analyzing the difference of features in related scales, a scale-space mapping algorithm is proposed to combine the similarity between the connected scales with the consistency in the same scale. With the quantization and evaluation of the segmentation results, the algorithm shows its property of stability, accuracy, generality and robustness to the noise disturbance.
出处
《自动化学报》
EI
CSCD
北大核心
2013年第10期1581-1593,共13页
Acta Automatica Sinica
基金
国家自然科学基金(60905005
61273237)
教育部博士点基金(20090111110015)
中央高校基本科研业务费专项资金(2012HGCX0001)资助~~
关键词
层次Markov随机场
集成标记
层间映射推理
图像分割
Hierarchical Markov random field (MRF), ensemble labeling, label-mapping inference, image segmentation