Otsu’s thresholding method is widely applied in image segmentation ,but it is not suitable to segment the images that have large difference in intra-class variance of object and background ,this paper presents a regularization Otsu’s segmentation method based on the information entropy of posterior probability of object and background obtained by thresholding image .From the point of view that image segmentation is essentially pixel clustering problem ,the classical Otsu’s thresholding method is firstly inter-preted as a kind of weighting hard C-means clustering algorithm .Secondly ,considering that the clustering segmentation has typical ill-posedness ,the object and background from raw image are obtained by segmentation and their posterior probability information en-tropy is taken as the constraint item ,the regularized modification of Otsu’s thresholding criteria function is realized ,and regulariza-tion Otsu’ s thresholding method is obtained .In the end ,the reasonability of the proposed thresholding method is explained by mathe-matical analysis ,and the selection method of its regularization parameter is put forward .Experimental results show that the proposed regularization Otsu’s thresholding method is effective ,and the traditional Ostu’ s thresholding method can be viewed as a special case of the proposed method .%Otsu阈值法是一种广泛应用的图像分割方法，但它并不适合目标和背景类内方差相差较大的图像分割需要，于是本文提出了一种基于分割所得目标和背景后验概率信息熵约束的正则化Otsu分割改进方法。从图像分割本质是像素聚类问题的观点出发，首先将传统Otsu阈值法解释为一种加权硬C-均值聚类；其次考虑到聚类问题具有典型的不适定性，利用图像分割所得目标和背景的后验概率所对应信息熵作为约束项，实现Otsu阈值化准则函数的正则化修改并得到了正则化Otsu阈值分割法；最后对该分割方法的合理性进行解释并给出其正则参数选取方法。实验结果表明，本文所建议的正则化Otsu阈值法是有效的，并将传统Otsu阈值法视为特例。
展开▼
