摘要
小波域隐Markov树模型(HiddenMarkovTreeModel,简称HMT)能充分表现小波系数的统计特征,但模型训练算法计算量大。文中以图像去噪为应用背景,提出了基于HMT粗分类的多树训练算法。该算法通过对不同类型的纹理建立不同的HMT,对小波系数进行粗分类,在此基础上,不同类别的小波系数被分别建模,并将粗分类HMT的参数作为最终模型训练的初始化参数,从而提高了模型的精度,同时减小了训练算法的计算量。对于含复杂场景或纹理的图像,提出了基于方差粗分类的训练算法,也能有效地提高模型精度。对自然图像和SAR图像的去噪实验表明,采用粗分类训练算法的HMT去噪模型的去噪效果在客观指标上优于现有的HMT去噪模型。
Wavelet domain hidden Markov tree (HMT) model can well capture the statistical characteristics of wavelet coefficients, but the computational complexity of its training algorithm is high. In this paper, we present a new training algorithm--HMT-based raw segmentation training algorithm for image denoising. By training different HMT models for different types of textures independently, the wavelet coefficient trees are separated into different types according to the textures' HMTs using a ML (maximum likelihood) classifier (raw segmentation). By this means, all trees of one type are modeled by one HMT which is initialized by the parameters of the raw segmentation HMT corresponding to the type of texture. For images containing complicated scenes or textures, we also present a simple block-variance-based raw segmentation algorithm which classifies the wavelet coefficient trees according to their corresponding block variance in the original image. Simulations on natural images and SAR (synthetic aperture radar) images indicate that the new algorithm performs better than the existing HMT-based algorithms in that its mean square error (MSE) is smaller and its computational overhead is less.
出处
《西北工业大学学报》
EI
CAS
CSCD
北大核心
2004年第4期457-462,共6页
Journal of Northwestern Polytechnical University
基金
国家自然科学基金 (60 3 750 0 3 )
航空基础科学基金 (0 3 I50 3 51 )资助
