对偶分布匹配的图像分割算法

摘要

Aiming at the weak points of local consistency measures,an image segmentation using dual distribution matching is proposed based on global consistency measures.The proposed algorithm uses the probability distribution of foreground and background as input distributions,and constructs dual matching model of foreground and background which describes the consistencies between the two input distributions and the resulting segmentation.Then the weighting parameters can be determined by utilizing the entire image distribution matching so that the global minimum of ε(L) captures the true solution L*.Finally,auxiliary functions of bhattacharyya measure graph cut is utilized to optimize ε(L) so that auxiliary labels La,Lb are alternately updated to converge to the true label L *.Experiments show that the segmentations are robust and accurate when input distributions are not so accurate.%针对传统局部一致性方法的缺点,在研究全局一致性方法的基础上,提出一种对偶分布匹配(Dual Distribution Matching,即DDM)的图像分割算法.该算法首先将前景和背景的概率分布作为输入分布,构造出前景和背景的对偶匹配模型,该模型描述两个输入分布和分割结果的相似度,然后利用整幅图像的分布来确定模型的权重参数,从而求解能量函数ε(L)的全局最小化的真解L*,最后利用基于Bhattacharyya的图分割(Bhattacharyya Measure Graph Cut,BMGC)的辅助函数完成能量函数ε(L)的优化,不断更新辅助标记La,Lb收敛于真实标记L*.实验表明在输入分布不够精确的情况下,该算法具有较好的准确性和稳定性.