Current Issue Cover

侯鑫, 陆耀, 李建武(北京理工大学计算机学院智能信息技术北京市重点实验室, 北京 100081)

摘 要
目的 非局部均值NLM(non-local means)方法是一种有效的数字图像去噪方法。然而,在实际去噪过程中,非局部均值的衰减参数通常是固定的而且无法随着图像的变化而作适应性的调整。为了使非局部均值方法更加有效, 提出一种将适用于多种噪声分布的无参考图像内容量度(表示为Q)引入NLM的迭代方法,来优化固定的衰减参数。方法 首先,针对普通图像的去噪,利用量度Q来测量每一次调整衰减参数后所对应的去噪结果的图像质量,凭借该迭代机制找到Q的最大值,从而获得最优的图像去噪结果;其次,将该量度用于MRI(magnetic resonance imaging)图像的去噪,利用Q来度量图像所含结构信息(如纹理和边缘),进而调整用于MRI图像去噪的无偏非局部均值法的衰减参数。结果 实验结果显示,本文方法提升了去噪结果的峰值信噪比(PSNR),并且本文方法的去噪结果在视觉上看起来比用传统方法得到的结果更清晰。结论 利用无参考图像内容量度Q来优化NLM方法的衰减参数,使得NLM方法能够针对不同的图像自适应地调整衰减参数以取得最优的去噪效果。实验结果表明用图像内容量度Q来优化非局部均值法的参数是有效的。
Optimizing the parameters of non-local means via no-reference image content metric for image denoising

Hou Xin, Lu Yao, Li Jianwu(Beijing Key Laboratory of Intelligent Information Technology, School of Computer Science and Technology, Beijing Institute of Technology, Beijing 100081, China)

Objective Nonlocal means (NLM) is a successful method to denoise digital images. However, the decay control parameter is usually fixed and cannot be adaptively changed according to the image content adaptively. Some intensified researches on the influence of the decay control parameter that describes the width of the smoothing kernel and data-dependence have been done to make the NLM more effective. Therefore, this paper proposes to apply the no-reference image content metric (denoted as Q) which does not need the noise to be Gaussian to NLM to optimize the fixed decay control parameter without a reference image. Method Some methods have been proposed to use SURE (Stein's unbiased risk estimate) as an estimator of mean squared error (MSE) of NLM to select the optimal parameters for restoring an image. However, these methods are fit only for Gaussian noise. Moreover, in real situations, the noise of an image is usually not Gaussian, and the variance of the noise is also space varying. Thus, in this situation, selecting the optimal parameters via SURE will be ineffective. Therefore, this paper first proposes to incorporate the no-reference image content metric Q which does not need the noise to be Gaussian to NLM to optimize the fixed decay control parameter in an iterative way and obtain the optimal denoised results. Second, aiming at the linear relationship between the Rician noise level in magnetic resonance imaging (MRI) images and the optimal decay control parameter, we propose to tune the parameter of unbiased NLM for MRI denoising according to the structure information (such as texture and edge) of the image. This method measures textures and edges in an image via the metric Q. By considering the structure information of the image, we can tune down the decay control parameter when the value of Q is high so that the textures and edges can be retained in the denoised results. Result Experimental results show that our proposed method can improve the value of peak signal-to-noise ratio. At the same time, the denoised results obtained by the proposed method are superior to the results obtained by the traditional methods in vision. Incorporating the image content metric to NLM is valid for denoising. Conclusion We propose to optimize the parameter of NLM via the no-reference image content metric Q. The proposed method aims to enable the NLM method to choose the decay control parameter adaptively according to the image content. Experimental results show that the proposed method can optimize the NLM in terms of vision and PSNR. Thus, incorporating the image content metric to NLM for denoising is valid.