信息化时代的发展使得数字图像成为信息的重要载体，从原始噪声图像恢复清晰图像也成为了几十年来的研究重点。自1992年ROF全变分模型的提出以来，全变分模型成为了图像去噪领域的研究重点，而全变分模型L1梯度项的存在会导致图像过于平滑，丢失掉图像原本的纹理细节信息，在边缘处产生严重的阶梯效应。全变分模型只考虑到了像素的局部邻域，没有把握全局信息，从而提出了非局部全变分模型，对每个图像块的相似度权重取值都考虑到了整个图像，更好地保留了图像的纹理信息，但是参数以及权重函数的取值在图像不同结构特征下的表现效果不同，影响整体复原效果。

本文我们主要介绍基于非局部全变分模型的两种改进算法：基于的改进模型将原本模型中的噪声图像替换成经过算法处理过的图像，这样的原始图像已经滤除了大部分噪声且保留了图像的纹理特征，减少了噪声对计算相似度权重的干扰，提升了模型的精确度，求解过程采用迭代算法也可以提升计算效率，通过计算机的不断迭代得到近似解，总体上提升了图像的去噪效果。基于的改进模型从视觉效果出发，将正则项的度量标准进行了转换，结合图像结构信息，能够在去噪的同时较好地保持图像整体视觉效果，同样采用迭代算法解决该最优化问题提升了模型整体的去噪能力，能够得到较满意的去噪图像。

With the development of the information age, digital images have become an important carrier of information. Restoring clear images from original noisy images has also become a research focus for decades. Since the ROF total variation model was proposed in 1992, the total variation model has become the focus of research in the field of image denoising. The existence of the L1 gradient term of the total variation model will cause the image to be too smooth, lose the original texture details of the image, and produce a serious staircase effect at the edge. The total variation model only considers the local neighborhood of the pixel and does not grasp the global information. Therefore, a non-local total variation model is proposed. The similarity weight value of each image block takes into account the entire image and better preserves the texture information of the image. However, the parameters and the weight function have different performance effects under different structural features of the image, which affects the overall restoration effect.

In this paper, we mainly study two improved algorithms based on non-local total variation model. The improved model replaces the noise image  in the original model with the image processed by the BM3D algorithm. This original image has filtered out most of the noise and retained the texture features of the image. It reduces the interference of noise on the calculation of similarity weights and improves the accuracy of the model. The iterative algorithm can also improve the computational efficiency. The approximate solution is obtained by continuous iteration of the computer, which improves the denoising effect of the image. Based on the visual effect, the improved model transforms the metric of the regularization term, combined with the image structure information, it can better maintain the overall visual effect of the image while denoising. The iterative algorithm is also used to solve the optimization problem, which improves the overall denoising ability of the model and can obtain satisfactory denoising images.