基于无噪图像块先验的MRI低秩分解去噪算法研究

Low Rank Decomposition for MRI Denoising based on Noise-free Image Patch Prior

针对核磁共振图像中存在莱斯噪声的现象,提出一种基于无噪图像块先验的MRI低秩分解去噪算法。该算法首先利用高斯混合模型学习无噪核磁共振图像块的先验;然后将带有无噪核磁共振图像块先验的高斯混合模型用于噪声核磁共振图像块聚类,并将聚类后每个高斯类中的核磁共振图像块叠在一起构成低秩矩阵并对其进行低秩分解操作来达到除去噪声的目的;最后根据去噪后的数据重建清晰核磁共振图像。实验结果表明相较于各项异性滤波,非局部均值滤波和权重核范数最小化复原算法,文中方法在PSNR值、SSIM值和视觉上有较大提升,在去除噪声的同时,能较好地保留图像本身的纹理细节信息。

In this paper, a low-rank matrix decomposition MRI denoising algorithm based on noise free image patch prior is proposed. Firstly, the algorithm learns the parameters of the Gaussian mixture model(GMM) from the noise-free MR image patch. The learned GMM with noise-free MR image patches priors is then used to guide the clustering of noisy MR image patches. Secondly, the image patch of noisy images in same Gaussian class are vectorized as a low-rank matrix. By a low-rank matrix decomposition process, the correspond denoised image data can be obtained. Thirdly, the clean image can be reconstructed from these denoised image data. Lastly, compared with the non-local means(NLM), the unbiased non-local means(UNLM),the anisotropic diffusion filtering(ADF) and the weighted nuclear norm minimization with variance stabilization transformation(WNNM-VST), our proposed method can effectively remove the Rician noise in the magnetic resonance image and has a great improvement in numerical results and visual effects.