Blog of Li

个人Blog 一年更新约一次

0%

每日论文: Data-driven tight frame construction and image denoising

我们知道,找到图像的稀疏表示可以在图像去噪、重建等方面有重大作用。常用的方法是字典学习,其中最广为人知的就是K-SVD方法。 但是,K-SVD方法的缺点主要在于其计算量大。DDTF的优点就在于计算量较小。DDTF模型来自Data-driven tight frame construction and image denoising

DDTF 模型和算法介绍

DDTF考虑如下的一个优化问题:

\[ \begin{equation} \begin{aligned}  \min_{v,W}& ||v-W(a_1,a_2,...a_m)g||_2^2+\lambda^2||v||_0\\ s.t.& \quad W^T W=I \end{aligned} \end{equation} \]

其中g是图像,g是编码,W是由m个filter,即\((a_1,a_2,...a_m)\)构成的紧框架。求解方法就是轮流最小化v和W。

  1. 给定\((W^{(k)})\),求解\((v^{(k)})\).

\[ \begin{equation} v^{(k)}:=argmin_v||v-W^{(k)}g||_2^2+\lambda^2||v||_0 \end{equation} \label{iterA} \]

  1. 给定\((v^{(k)})\),求解\((W^{(k+1)})\).

\[ \begin{equation} \{a_i^{(k+1)}\}_{i=1}^m:=argmin_{\{a_i\}_{i=1}^m}||v^{(k)}-W(a_1,a_2,...a_m)g||_2^2 \end{equation} \label{iterB} \]

对于(),存在唯一的最优解\(( {\bar v}=T_{\lambda}(Wg))\), 对于(),可以通过SVD算法精确求解:

假设filter大小为\((r \times r)\),则共有\((r^2)\)个filter。 Here \((g_n, n=1,2,...N)\) are all \((r\times r)\) patches from the input image g.

Let

\[ {\vec v}_n=(v^{(k),1}(n),v^{(k),2}(n),...,v^{(k),r^2}(n))^{T}, 1 \leqslant n \leqslant N. \] and define

\[ \begin{equation} \left\{ \begin{aligned}  V=({\vec v}_1,{\vec v}_2,...{\vec v}_N) \in \mathbb{R}^{r^2\times N}\\  G=({\vec g}_1,{\vec g}_2,...{\vec g}_N) \in \mathbb{R}^{r^2\times N}\\  A=({\vec a}_1,{\vec a}_2,...{\vec a}_N) \in \mathbb{R}^{r^2\times r^2} \end{aligned} \right. \end{equation} \]

Then we have

\[ \begin{equation} \begin{aligned}  ||v-Wg||_2^2&=\sum_{n=1}^N||{\vec v}_n-A^T{\vec g}_n||_2^2\\ &=\sum_{n=1}^N {\vec v}_n^T {\vec v}_n + {\vec g}_n^T A A^T {\vec g_n} - 2{\vec v}_n^T A^T {\vec g}_n \\ &= \sum_{n=1}^N {\vec v}_n^T {\vec v}_n + \frac{1}{r^2} {\vec g}_n^T{\vec g_n} - 2(A{\vec v}_n)^T{\vec g}_n \\ &= Tr(V^T)+\frac{1}{r^2}Tr(G^T G)-2Tr(AVG^T) \end{aligned} \end{equation} \]

所以对于(),等价于:

\[ \begin{equation} \max_A Tr(AVG^T) \quad s.t. \quad A^TA=\frac{1}{r^2}I_{r^2} \end{equation} \]

引理1 Let B and C be \((m\times r)\) matrices where B has rank r. Consider the constrained maximization problem:

\[ B_*=\arg\min_B Tr(B^T C) \quad s.t. \quad B^TB=I_r \] Suppose that the single value decomposition (SVD) of \((C)\) is \(C = UDX^T\).Then \(( B_*=UX^T )\).

根据引理,我们就能知道,通过SVD计算就可以求得()的精确解。

讨论

与K-SVD的主要区别在于通过加上了紧框架的约束, 反而构造出了一个相对容易求解的优化模型。但同时也失去了冗余字典模型的优势。

与 K-SVD 的异同

Method 是否来自数据 是否正交 计算方法 计算速度
K-SVD 冗余字典 每步为近似求解
DDTF 正交字典 每步为较精确求解
Wavelets 都有 直接构造

正交字典的优劣

来自: Orthogonal Matching Pursuit and K-SVD for Sparse Encoding ### Benefits of ONB

  • Analytic formulations
  • Well understood mathematical properties
  • Fast algorithms for projection

Limitations

  • Orthonormal bases are optimal only for specific synthetic signals
    • If your signal looks exactly like your basis, you only need one coefficient
  • Limited expressiveness, all signals behave the same
  • Real world signals often take a lot of coefficients
    • Just truncate the series, which leads to artifacts like aliasing

参考文献:

Data-driven tight frame construction and image denoising

Orthogonal Matching Pursuit and K-SVD for Sparse Encoding