复小波包变换域混合统计模型图像降噪算法  

Image denoising algorithm using mixed statistical model in complex wavelet packet transform

作　　者：闫河[1] 何光敏 张小川[1] 

机构地区：[1]重庆理工大学计算机学院,重庆400054 [2]重庆市天宝实验学校,重庆400050

出　　处：《控制理论与应用》2010年第3期335-343,共9页Control Theory & Applications

基　　金：国家自然科学基金资助项目(60443004);重庆市科委自然科学基金资助项目(CSTC;2008BB2340);重庆市教委科学技术研究项目(KJ080621);重庆理工大学科研启动基金项目(2009ZD12)

摘　　要：该方法利用四树复小波包变换具有的移不变性、良好的方向选择性和对高频信号的细致分析能力等特点,把含噪图像分解成低频逼近子图和若干高频方向子图;在保留低频逼近子图复系数不变的同时,利用复系数层间相关性的强弱把高频方向子图分为主要类和次要类.对主要类和次要类复系数分别进一步采用非高斯双变量模型和零均值高斯分布模型进行噪声抑制.实验结果表明,无论是峰值信噪比(PSNR)指标,还是在视觉效果上,本文方法的去噪性能均好于传统的双树复小波变换去噪、四树复小波包变换去噪和小波域高斯尺度混合模型去噪,在有效抑制噪声的同时,具有很好的图像边缘和细节保护能力.The noisy image is decomposed into low frequency approximate subimages and high frequency directional subimages by using the quad-tree complex wavelet packet transforrn（QCWPT） which has the advantages of shift-invariance, high directional resolution and fine discrimination of high frequency signals. The complex coefficients in low frequency approximate subimages are kept unchanged, while the high frequency directional subimages are categorized as major type and minor type according to their inter-scale correlation. Noises in both types are removed by using of the non-Gaussian bivariate model and the zero mean Gaussian distributing model, respectively. In comparing either the power signal-to-noise ratio（PSNR） index or the visual effects with other methods, the presented scheme outperforms the traditional dual-tree complex wavelet transform, QCWPT and wavelet domain Gaussian scale mixtures. Experiments also show that the presented scheme achieves an excellent balance between the suppression of noises and the preservation of image details and edge.

关 键 词：图像去噪 四树复小波包变换 层间相关性 非高斯双变量模型:零均值高斯分布模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]