基于小波-双马尔可夫随机场的纹理分割

Texture segmentation based on wavelet-double Markov random fields

传统高斯金字塔马尔可夫随机场模型仅仅捕捉图像低频信息,纹理分割效果不太理想.根据纹理图像小波分解后各频带的统计性质和层次关系,优化频带选取,提出了一种变形小波结构,建立了融合这种结构上尺度内部和尺度之间关系的双马尔可夫随机场模型,引入了一种近似最大联合概率分割算法,并从理论上分析了该算法的合理性.实验表明,与基于高斯金字塔马尔可夫随机场模型的分割方法相比,该算法分割质量明显提高;并且,对模型中自由参数的选取进行比较,证实它们在给定区间上的选择具有鲁棒性.

Traditional Gauss-pyramid markov random fields only catch information in low frequency channels, so rex ture segmentation results are not satisfying. Due to some special statistical properties in channels and some relations between layers after wavelet decomposition, a deformed wavelet structure is presented by selecting appropriate chan nels, then double markov random fields based on intralscale and interscale relations is built and an approximate-maximal-joint-probability algorithm for texture segmentation is produced, This algorithm is proved accurate in terms of theoretical analysis. Compared with the segmentation based on Gauss pyramid markov fields, the experiments show that the proposed approach has superior results. Furthermore, some parameters are robust independent of their selection in some given intervals.