Improving Nonlocal Means Method via a No-reference Image Content Metric for MRI Denoising  

作　　者：HOU Xin LI Jianwu LU Yao DONG Zhengchao 

机构地区：[1]Beijing Key Laboratory of Intelligent Information Technology, School of Computer Science and Technology,Beijing Institute of Technology [2]Brain Imaging Lab, Columbia University [3]New York State Psychiatric Institute

出　　处：《Chinese Journal of Electronics》2014年第4期735-741,共7页电子学报（英文版）

基　　金：supported by the National Natural Science Foundation of China(No.61271374,No.61273273);the Beijing Natural Science Foundation(No.4122068)

摘　　要：The patch matching of the traditional Nonlocal means （NLM） filter mainly depends on structure similarity and cannot adapt to the patch rotation or mirroring transformation. Therefore, designing a measure with rorationally invariant similarity is of significant importance for improving the effectiveness of patch comparison of NLM. This paper proposes to apply a no-reference image content metric with the rotation-invariance to NLM for denoising Magnetic resonance （MR） images. The metric measures quantitatively the content of a patch in an image, including sharpness, contrast, and geometric features such as textures and edges. The metric values for every patch are computed and added into the Gaussian matching kernel of NLM so as to effectively perform patch matching. The main advantage of the proposed method is that it does not need to rotate patches in different orientations during patch matching. Experimental results show that the proposed method is superior to the traditional NLM, the state-of-the-art method Block-matching and 3- D （BM3D） filtering and the Unbiased NLM （UNLM） for MRI denoislng.The patch matching of the traditional Nonlocal means(NLM) filter mainly depends on structure similarity and cannot adapt to the patch rotation or mirroring transformation. Therefore, designing a measure with rotationally invariant similarity is of significant importance for improving the effectiveness of patch comparison of NLM. This paper proposes to apply a no-reference image content metric with the rotation-invariance to NLM for denoising Magnetic resonance(MR) images. The metric measures quantitatively the content of a patch in an image, including sharpness, contrast, and geometric features such as textures and edges. The metric values for every patch are computed and added into the Gaussian matching kernel of NLM so as to effectively perform patch matching. The main advantage of the proposed method is that it does not need to rotate patches in different orientations during patch matching. Experimental results show that the proposed method is superior to the traditional NLM, the state-of-the-art method Block-matching and 3-D(BM3D) filtering and the Unbiased NLM(UNLM) for MRI denoising.

关 键 词：Nonlocal means Rotationally invariant similarity measure No-reference image content metric Magnetic resonance image denoising. 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TQ172.622[自动化与计算机技术—计算机科学与技术]