【摘要】 本文提出了一种基于分布特征的多尺度无监督图像分割方法。通过对多尺度图像数据在每个尺度上进行Gauss子集聚类,并将每个像素的邻域内的Gauss子集类别标记作为特证向量,利用多尺度Markov模型进行二次聚类,从而实现无监督图像分割。与其它基于多尺度Markov模型的无监督分割方法和传统动态聚类方法相比,该方法既无需假定每类的分布形式,又能较好地反映数据的概率结构。对合成图像与SAR图像的实验结果表明,该方法的分割精度接近于有监督的H-MPM和H-SMAP方法。更多还原

【Abstract】 A new multiscale Markov model based Bayesian approach to image segmentation is presented. By Gauss mixture model and MAP estimation, the image data are first clustered into different Gauss classes. Then by modeling the Gauss class labels with Markov quadtree and MPM estimation, the final segmentation is performed. Compare with existing continuous version segmentation algorithm based on multiscale Markov model, the new approach needn’t assuming the distribution form of each class known. And compared with existing discrete version segmenta- tion method, the feature data in our approach take very limit values, and so the number of distribution parameters is small. Moreover, because the feature data are based on neighborhood, the segmentation can be more smoothed, and the estimation uncertainty can be reduced. Experimental results show that the unsupervised approach can give a com- parable segmentation with supervised H-MPM and H-SMAP.更多还原