摘要
针对两个状态的有限高斯混合模型逼近小波系数的不足和小波域隐马尔可夫树标号场相互独立的缺点,提出了一种基于小波域层次马尔可夫模型的图像分割算法,这种模型用有限通用混合模型逼近小波系数的分布,使有限高斯混合模型只是其一种特殊情况;在标号场的先验模型确定上,利用马尔可夫模型描述标号场的局部作用关系,给出标号场的具体表达式,克服了小波域马尔可夫树模型标号场相互独立的不足,然后利用贝叶斯准则,给出相应的分割因果算法。该模型不仅具有空域马尔可夫模型有效的递归算法的优点,同时具有小波域隐马尔可夫树模型中的马尔可夫参数变尺度行为。最后用真实的图像和合成图像同几种分割方法进行了对比实验,实验结果表明了本文算法的有效性和优异性。
In order to overcome the deficiency of approximation to the wavelet coefficient joint probability with two-state Gaussian mixture model（GMM） and the shortcoming of the independence between wavelet labels in wavelet domain hidden Markov tree model（ HMT）, a new image segmentation algorithm based on wavelet domain hierarchical Markov model is proposed. The new image model is described as wavelet coefficient joint distribution with finite general mixture model （ FGM ）, while the GMM in HMT model is only one of the FGMs. Vitilizing on the local interactions of labels described by Markov random field （MRF）, the label field priori probability mode] with explicit expression, which overcomes the shortcoming of the independence between labels in the HMT mode], is determined. Using Bayes principle, the recursive algorithm of image segmentation is derived. The proposed model inherits not only the characteristics of spatial domain hierarchical MRF model with effective recursive algorithm but also the characteristics of HMT model with the variable Markov parameters in different scales. The experiments with real images and synthetic texture images are carried out, the results show that the proposed method outperforms other standard segmentation methods, such as accurately locating image edges, correctly identifying different regions.
出处
《中国图象图形学报》
CSCD
北大核心
2007年第2期308-314,共7页
Journal of Image and Graphics
关键词
小波域马尔可夫随机场
最大后验概率
图像分割
EM算法
wavelet domain Markov random field, maximum a posterior （MAP） probability, image segmentation,Expectation-maximization algorithm
