维纳滤波图像去噪_BM3D 图像去噪算法||原论文翻译

偶然看到某知乎大佬Uno Whoiam对BM3D的翻译，其翻译地址如下，可能大佬比较忙，没没翻译完整，我就试着狗尾续貂，把剩下的算法部分翻译完整，不足之处请各位大佬批评指正，蟹蟹！！

论文前面部分的翻译：

Uno Whoiam：BM3D 图像去噪算法||原论文翻译 1/N​zhuanlan.zhihu.com

D. Collaborative filtering by shrinkage in transform domain

An effective collaborative filtering can be realized as shrinkage in transform domain. Assuming d+1-dimensional groups of similar signal fragments are already formed, the collaborative shrinkage comprises of the following steps.

(1) Apply a d+1-dimensional linear transform to the group.

(2) Shrink (e.g. by soft- and hard-thresholding or Wiener filtering) the transform coefficients to attenuate the noise.

(3) Invert the linear transform to produce estimates of all grouped fragments.

D. 在变换域通过收缩进行协同滤波

一个有效的协同滤波可以通过在变换域进行收缩来实现。假设已经形成了d+1维的由相似信号片构成的分组，则协同滤波包括以下的步骤：

(1)应用一个d+1维的线性变换到组；

(2) 收缩（例如软/硬阈值或者维纳滤波）变换系数以减弱噪声；

(3) 执行反变换获得所有分组片段的估计。

This collaborative transform-domain shrinkage can be particularly effective when applied to groups of natural image fragments, e.g. the ones in Figure 1. These groups are characterized by both:

（1）intra-fragment correlation which appears between the pixels of each grouped fragment - a peculiarity of natural images;

（2）inter-fragment correlation which appears between the corresponding pixels of different fragments - a result of the similarity between grouped fragments。

这个协同变换域收缩在应用到自然图像片段（如图1的那些片段）构成的分组时特别有效。这些分组可以描述为两个特征：

（1）每一个已经分组的片段内像素间的相关性-自然图像的特性；

（2）不同片段之间对应位置的像素间相关性-分组片段间的相似结果。

The 3D transform can take advantage of both kinds of correlation and thus produce a sparse representation of the true signal in the group. This sparsity makes the shrinkage very effective in attenuating the noise while preserving the features of the signal.

Let us give a simple illustration of the benefit of this collaborative shrinkage by considering the grouped image blocks shown in Figure 1. Let us first consider the case when no collaborative filtering is performed but instead a 2D transform is applied separately to each individual block in a given group of n fragments. Since these grouped blocks are very similar, for any of them we should get approximately the same number, say

, of significant transform coefficients. It means that the whole group of n fragments is represented by

coefficients. In contrast, in the case of collaborative filtering, in addition to the 2D transform, we apply a 1D transform across the grouped blocks (equivalent to applying a separable 3D transform to the whole group). If this 1D transform has a DC-basis element, then because of the high similarity between the blocks, there are approximately

only

significant coefficients that represent the whole group instead of

. Hence, the grouping enhances the sparsity, which increases with the number of grouped blocks.

3维变换可以使用上述两种相关性并且在组中生成一个真实信号的稀疏表示。这个稀疏表示在保留信号特征的同时使得通过收缩去减弱噪声是有效的。

让我们用图1中已经分组的图像块来举一个简单的例子说明协同收缩（其实就是协同滤波时使用的收缩算法）的好处。我们首先考虑没有使用协同滤波而是将2维变换分开地应用到包含n个片段的每一个单独的块的情形，因为这些块（译者注：也就是片段）是十分相似的，对于它们中的每一个，我们应该得到一个近似相同的数字，比如说是

,一个很重要的变换系数。这就意味着这n个片段组成的组由

个系数来表示。相比之下，在使用协同滤波的情形，除了2维变换，我们还应用一个1维的变换到已经分组的块（等价于应用一个可分离的3维变换到整个组）。如果这个1维变换有DC-basis（这里我认为是离散余弦基）成分的话，那么因为块之间的高度相似性，使得几乎只有

个重要的系数来表示整个组而不是

个。这样子的话，分组加强了稀疏性，并且稀疏性随着分组块数量的增加而加强。

注解1：This is just a qualitative statement because the actual number of significant coefficients depends on the normalization of the transforms and on the thresholds used for the 2D and 3D cases.

这仅仅是一个定型的陈述，因为重要系数的实际数目取决于变换的标准化和所使用的二维和三维的阈值。

As Figure 1 demonstrates,a strong similarity between small image blocks at different spati