Image denoising based on nonconvex anisotropic total-variation regularization
Image denoising based on nonconvex anisotropic total-variation regularization
复制标题
DOI：
10.1016/j.sigpro.2021.108124
复制
发表时间：
2021-04
期刊：
影响因子：
--
通讯作者：
Junchen Guo;Qinghua Chen
中科院分区：
文献类型：
--
作者：
Junchen Guo;Qinghua Chen
基于全变分正则化的图像去噪模型在图像处理的许多领域得到了广泛的应用。TV正则化的主要优点是在恢复退化图像时可以有效地保持边缘。但是，由于TV项中梯度子变量的权重相同，其主要缺陷是不能描述局部特征。为此，我们提出了一个非凸各向异性全变分（NCATV）的图像去噪模型。在该模型中，加权矩阵依赖于恢复图像，所以我们可以期望它可以描述局部特征更强大。由于所提出的模型的非凸性，首先，我们需要使用逐次替换方案与加权矩阵从TV项解耦。通过这种操作，所提出的模型转化为非光滑和凸优化问题。然后利用交替方向乘子法（ADMM）将该问题转化为几个易于求解的子问题。数值实验表明，我们提出的模型产生了改善性能的视觉和定量相比，一些国家的最先进的方法。
Image denoising models based on the total variation (TV) regularization have been used in many fields of image processing. The main advantage of the TV regularization can preserve the edge efficiently when restoring the degraded image. However, the main flaw can not describe the local feature due to the same weight for the gradient subvariable in the TV term. To this end, we propose a nonconvex anisotropic total variation (NCATV)-based image denoising model. In the proposed model, the weighted matrix depends on the restored image, so we can expect that it can describe local features to be more robust. Due to the nonconvexity of the proposed model, first we need to use the successive replacement scheme to decouple with the weighted matrix from the TV term. With this operation, the proposed model is transformed into the nonsmooth and convex optimization problem. Then we can employ the alternating direction method of multipliers (ADMM) to transform this problem into several easily solvable subproblems. Numerical experiments show that our proposed model yields an improvement in performance both visually and quantitatively compared with some state-of-the-art methods.