基于移不变抗混叠轮廓波变换的混合统计模型图像降噪

Mixed statistical model image denoising based on shift-invariant non-aliasing Contourlet transform

针对抗混叠轮廓波变换缺乏平移不变性的缺陷,构造出具有近似移不变性的抗混叠轮廓波变换。在此基础上,在变换域提出一种混合统计模型图像降噪方法。该方法充分利用变换域信号系数层间层内相关性强、噪声系数无层内相关性且在小尺度下存在较强的假层间相关性的特点,采用混合统计模型对小尺度信号系数进行估计,从而避免了非高斯双变量模型放大噪声系数的风险。实验结果表明,提出的去噪法能克服轮廓波变换中的频谱混叠,避免重构图像出现"划痕"和边缘模糊现象,得到的峰值信噪比(PSNR)值分别比轮廓波硬阈值去噪、轮廓波变换域HMT去噪和抗混叠轮廓波变换域硬阈值去噪平均高2.87,1.32和1.36 dB,在有效去噪的同时,具有较好的图像边缘和细节保护能力。

To avoid shift-variance defects in the original Non-aliasing Contourlet Transform （NACT）, a new approximate Shift-invariance NACT（SINACT） was proposed. On this basis, a mixed statistical model image denoising method was presented based on SINACT. This method took full advantage of the characteristics that there were intra-scale and inter-scale correlations for signal coefficients and there was no intra-scale correlation but strong inter-scale correlation for noise coefficients at small scales. Furthermore, a mixed statistical model was used to estimate the small-scale signal coefficients to avoid noise coefficients amplified by the non Gaussian bivariate model. Experimental results show that the proposed scheme can overcome the aliasing in the Contourlet transform domain and can avoid ＂scratching＂ and edge blur phenomena in the reconstructed image. The denoising Peak Signal to Noise Ratio（PSNR） of the proposed scheme is on average higher by about 2.87,1.32 and 1.36 dB than those of the Contourlet transform hard-threshold denoising, Contourlet transform domain HMT denoising and hard-threshold denoising based on NACT, respectively,and it can achieve an excellent balance between suppressing noise and preserving as many image details and edges as possible.