小波变换估计非线性扩散最优停止时间

Gilboa提出一种针对高斯噪声的基于信噪比(SNR)最优的迭代停止时间估计方法。该方法用一个噪声补丁来估计图像噪声与冗余(噪声图像与去噪图像的差)的协方差对冗余方差的导数,补丁是随机生成的纯高斯噪声图像,其均值为零并且方差等于噪声图像的噪声方差。在实际应用中图像噪声方差未知,补丁的噪声是随机的,不同噪声所得到的最后停止时间可能不同。针对这些问题,对该方法进行了改进。首先将图像进行小波变换;再利用小波系数的层间相关性去掉第1层斜向高频系数(HH1)中的边缘纹理信息,获得"纯"的子噪声;然后把子噪声作为补丁的噪声取代随机噪声。实验结果表明,改进方法不仅能解决随机噪声补丁的两个问题,而且去噪图像在峰值信噪比(PSNR)上有一定优势。

Stopping-time estimation for anisotropic diffusion using discrete wavelet transform

Anisotropic diffusion(ATD) is a very important method for image denoising.The selection of the optimal stopping-time for ATD is one of the most important problems. Recently, Gilboa proposed an estimation method of stopping-time for ATD in Gaussian noisy images based on an optimal SNR. The method uses a noisy patch to estimate the derivative of the covariance of the noise and the redundancy (the result of noisy image minus the denoised image) with respect to the variance of the redundancy. The patch's noise is random Gaussian noise whose mean is zero and whose variance is the variance of the image's noise. The method has two defects. On the one hand, the method needs the variance of the image's noise,which is unknown in practice. On the other hand, the patch's noise is random and the result may be different because of different patch's noise. Our proposed method is optimized for these problems. First, the noisy image is transformed by wavelets. Then the information of edges and textures in the first coefficients of direct wavelet (HH₁) is reduced by using the inter-scale correlation of wavelet coefficients. Last, the reduced HH₁ is taken as the patch's noise. Experiments show that the proposed method can solve the two defects and the denoised image by the proposed method has a better PSNR.