多尺度Markov模型的可适应图像分割方法

A Method of Adaptive Image Segmentation Based on Multiscale Markov Models

本文在图像分割的TSMAP(trainablesequentialmaximumaposterior)方法基础上,提出基于多尺度Markov模型的可适应ATSMAP(adaptiveTSMAP)图像分割方法.在给定训练图像及其基本真实分割(groundtruthsegmentation,GTS)的基础上,通过直接对原始图像的GTS进行小波变换产生粗尺度上的GTS,进而估计出图像数据的分布参数和Markov四叉树模型参数;上下文模型参数根据上下文的低维特征(类别数量特征)而非上下文本身来估计.该方法具有上下文模型参数估计计算量小,Markov四叉树模型参数可针对特定的待分割图像重新优化等优点(模型适应过程),解决了TSMAP方法易导致过学习的问题,在待分割图像与训练图像的统计特性不匹配的情况下,仍能给出较好的分割结果.对合成图像与SAR图像的实验结果表明,这种方法的分割精度高于TSMAP和其它几种基于多尺度Markov模型的图像分割方法.

In order to remedy some of the disadvantages of trainable sequential maximum a posterior （TSMAP）, we introduce an adaptive trainable sequential maximum a posterior （ATSMAP） approach for image segmentation based on multiscale Markov model. This method obtains the ground truth segmentations at all scales by wavelet decomposition of the accurate segmentation of the original image, and then estimates the parameters of mulfiscale quadtree models. Also, the parameters of the context model are estimated by a simple way that estimates the parameters with the low dimension feature instead of the context itself. Compared with TSMAP, the advantages of ATSMAP are as follows：1）.The parameters of quadtree can be retrained for the specific image being segmented.2） The estimation of parameters for context models is very computationally efficient. The experimental results of the segmentation for both synthetic images and SAR images indicate that the approach fairly improves the segmentation accuracy over TSMAP and some other multiscale Markov based methods.