Paper Reading on Deep Compressed Sensing

2019.09.22
ludics

1. Learning a Compressed Sensing Measurement Matrix via Gradient Unrolling

1.1 Abstract

Linear encoding of sparse vectors is widely popular, but is commonly data-independent – missing any possible extra (but a priori unknown) structure beyond sparsity. In this paper, we present a new method to learn linear encoders that adapt to data, while still performing well with the widely used l1\mathcal{l}_1 decoder. The convex l1\mathcal{l}_1 decoder prevents gradient propagation as needed in standard gradient-based training. Our method is based on the insight that unrolling the convex decoder into T projected subgradient steps can address this issue. Our method can be seen as a data-driven way to learn a compressed sensing measurement matrix. We compare the empirical performance of 10 algorithms over 6 sparse datasets (3 synthetic and 3 real). Our experiments show that there is indeed additional structure beyond sparsity in the real datasets; our method is able to discover it and exploit it to create excellent reconstructions with fewer measurements (by a factor of 1.1-3x) compared to the previous state-of-the-art methods. We illustrate an application of our method in learning label embeddings for extreme multi-label classification, and empirically show that our method is able to match or outperform the precision scores of SLEEC, which is one of the state-of-the-art embedding-based approaches.

1.2 Introduction

Design a measurement matrix ARm×dA \in \mathbb{R}^{m \times d}, such that for an ill-posed problem y=Axy = Ax where xRdx \in \mathbb{R}^d and yRmy \in \mathbb{R}^m, and m<dm < d. The problem of designing measurement matrices and reconstruction algorithms that recover sparse vectors from linear observations is called Compressed Sensing (CS), Sparse Approximation or Sparse Recovery Theory.

The problem can be solved as follows:

arg minxRdx0s.t.Ax=y \argmin _{x' \in \mathbb{R}^d} \parallel x' \parallel _0 \quad \mathrm{s.t.} \quad Ax'=y

But it's a NP-Hard problem. With some properties such as Restricted Isometry Property (RIP) or the nullspace condition (NSP), l0\mathcal{l}_0 norm can be relaxed to l1\mathcal{l}_1 norm:

D(A,y)=arg minxRdx1s.t.Ax=y D(A, y) = \argmin _{x' \in \mathbb{R}^d} \parallel x' \parallel _1 \quad \mathrm{s.t.} \quad Ax'=y

The authors interested in vectors that are not only sparse but have additional structure in their support. They propose a data-driven algorithm that learns a good linear measurement matrix AA from data samples. Their linear measurements are subsequently decoded with the l1\mathcal{l}_1-minimization to estimate the unknown vector x.

Our method is an autoencoder for sparse data, with a linear encoder (the measurement matrix) and a complex non-linear decoder that approximately solves an optimization problem.

PCA is a data-driven dimensionality reduction method and an autoencoder with encoder and decoder all linear, and is provides the lowest MSE. Linear encoder has advantages, 1) easy to compute matrix-vector multiplication; 2) easy to interpret as every column of the encoding matrix can be viewed as feature embedding. Arora et al. found that pre-trained word embeddings such as GloVe and word2vec formed a good measurement matrices for text data, which need fewer mearusements than random matrces when used with l1\mathcal{l}_1-minimization.

Given nn sparse samples x1,x2,,xnRdx_1, x_2, \dots, x_n \in \mathbb{R}^d, our problem is to find the best AA:

minARm×df(A),wheref(A):=i=1nxiD(A,Axi)22 \min \limits_{A \in \mathbb{R}^{m \times d}} f(A), \quad \mathrm{where} \quad f(A) := \sum_{i=1}^{n} \parallel x_i - D(A, Ax_i) \parallel_2^2

where D(,)D(·,·) is the l1\mathcal{l}_1 decoder.

f(A)f(A)关于AA的梯度很难求解,因为D(,)D(\cdot, \cdot)是由一个优化问题确定的。作者的创新点就在这里,将l1\mathcal{l}_1最小化转变成了T步的投影子梯度的更新,这样梯度就可以计算了。即

f^(A):=i=1nxix^i22,wherex^i=TstepsubgradientofD(A,Axi),fori=1,2,,n. \begin{aligned} \hat{f}(A) :&= \sum_{i=1}^n \lVert x_i - \hat{x}_i\rVert^2_2, \mathrm{where} \\ \hat{x}_i &=T\mathrm{step \, subgradient \, of\, } D(A,Ax_i), \mathrm{for}\, i = 1,2,\cdots,n. \end{aligned}

这个步骤就被称为梯度展开(gradient unrolling)。推导过程:

minxRdx1s.t.Ax=y \min_{x'\in \mathbb{R}^d} \lVert x' \rVert_1 \quad \mathrm{s.t.} \quad Ax'=y
x(t+1)=Π(x(t)αtg(t)),whereg(t)=sign(x(t)) x^{(t+1)} = \Pi(x^{(t)} - \alpha_t g^{(t)}), \quad \mathrm{where} \quad g^{(t)} = \mathrm{sign}(x^{(t)})

其中,Π\Pi表示到凸集{x:Ax=y}\{x':Ax'=y\}上的投影,g(t)g^{(t)}是符号函数,也就是1\lVert\cdot\rVert_1x(t)x^{(t)}处的子梯度,αt\alpha_t是第t个迭代的步长。因为AA是行满秩矩阵,Π\Pi有如下闭式解:

Π(z)=arg minhhz22s.t.Ah=y=z+arg minhh22s.t.Ah=yAz=z+A+(yAz) \begin{aligned} \Pi(z) & = \argmin_h\lVert h-z\rVert_2^2 \quad \mathrm{s.t.} Ah=y \\ &=z + \argmin_{h'}\lVert h'\rVert_2^2 \quad \mathrm{s.t.} Ah'=y-Az \\ &= z + A^{+}(y-Az) \end{aligned}

这里的A+=AT(AAT)1A^+ = A^\mathrm{T}(AA^\mathrm{T})^{-1}AA的伪逆。带入上式,又Ax(t)=yAx^{(t)}=y,有

x(t+1)=x(t)αt(IA+A)sign(x(t)) x^{(t+1)} = x^{(t)} - \alpha_t (I-A^+A)\mathrm{sign}(x^{(t)})

使用x(1)=A+yx^{(1)}=A^+y作为起始点。

根据一个引理,我们有理由使用ATA^\mathrm{T}来替换A+A^+,因为A+A^+的计算很麻烦,而且难以反向传播。具体的引理内容参见论文。最终的迭代形式成为:

x(t+1)=x(t)αt(IATA)sign(x(t)) x^{(t+1)} = x^{(t)} - \alpha_t (I-A^\mathrm{T}A)\mathrm{sign}(x^{(t)})

引理要求AA的奇异值都为1,训练时并没有添加这个约束,但是最终得到的测量矩阵和约束集很接近。网络结构如下:

Net Arch
Net Arch

经过TTblock,输出为x(T+1)x^{(T+1)},再进入激活层,最终输出为x^=ReLU(x(T+1))\hat{x}=\mathrm{ReLU}(x^{(T+1)})。对于给定的n个无标签训练数据,我们训练的目的是减小xxx^\hat{x}之间的均方l2\mathcal{l}_2距离:

minARm×d,βR1ni=1nxx^22 \min_{A\in \mathbb{R}^{m\times d}, \beta\in \mathbb{R}}\frac{1}{n}\sum_{i=1}^n \lVert x-\hat{x}\rVert_2^2

对于这样学到的线性编码器AA,我们可以使用l1\mathcal{l}_1解码器进行解码,作者使用Gurobi进行解码,因此整套算法可以写成L1-AE+L1-min pos。这样学到的编码器与l1\mathcal{l}_1解码器和基于模型的解码器都可以有很好的表现。

作者还发现L1AEL1AE可以用于XML(极多标签学习)。

2. Deep Compressed Sensing

2.1 Abstract

Compressed sensing (CS) provides an elegant framework for recovering sparse signals from compressed measurements. For example, CS can exploit the structure of natural images and recover an image from only a few random measurements. CS is flexible and data efficient, but its application has been restricted by the strong assumption of sparsity and costly reconstruction process. A recent approach that combines CS with neural network generators has removed the constraint of sparsity, but reconstruction remains slow. Here we propose a novel framework that significantly improves both the performance and speed of signal recovery by jointly training a generator and the optimisation process for reconstruction via metalearning. We explore training the measurements with different objectives, and derive a family of models based on minimising measurement errors. We show that Generative Adversarial Nets (GANs) can be viewed as a special case in this family of models. Borrowing insights from the CS perspective, we develop a novel way of improving GANs using gradient information from the discriminator.

压缩传感(CS)提供了一个优雅的框架,可从压缩测量中恢复稀疏信号。例如,CS可以利用自然图像的结构并仅从几个随机测量中恢复图像。CS具有灵活性和数据效率,但是其应用受到稀疏性的强烈假设与昂贵的重建过程的限制。将CS与神经网络生成器结合在一起的最新方法消除了稀疏性的约束,但是重建仍然很慢。在这里,我们提出了一个新颖的框架,该框架通过联合训练生成器和通过元学习进行重构的优化过程,可以显着提高信号恢复的性能和速度。我们探索了训练具有不同目标的测量,并基于最小化测量误差得出一系列模型。我们表明,在该系列模型中,可以将生成对抗网络(GAN)视为特例。从CS角度汲取见解,我们开发了一种使用鉴别器中梯度信息来改进GAN的新颖方法。

2.2 Introduction

编解码是通信的核心问题,压缩感知(CS)提供了一个将编码和解码分离成独立的测量和重建过程的框架。

但是CS在处理大规模数据时却因为稀疏假设与重建效率而被阻碍。Bora et al., 2017将CS与独立训练的神经网络生成器合并。本文则提出深度压缩感知(Deep Compressed Sensing,DCS)框架,神经网络同时对测量和重建从头训练。这个框架自然导出了一系列模型,包括GAN,是通过训练不同目标的测量函数得到。文章的贡献有:

2.3 Background

2.3.1 Compressed Sensing

参考自Donoho, 2006; Candes et al., 2006

m=Fx+η \boldsymbol m = \mathbf{F}\boldsymbol x + \eta

式中η\eta是高斯噪声,FFC×DC\times D的测量矩阵,CDC\ll D,这样mm的维度远低于xx。当xx是稀疏,且FF是随机矩阵时,CS可以以很高的概率完美恢复xx。实践中,xx的稀疏性要求可以被xx在某个基Φ\Phi下的稀疏性所替代,如傅立叶基或小波基,这样Φx\Phi x可以表示非稀疏信号如自然图像等。CS理论的核心是限制保距性(Restricted Isometry Property,RIP),其定义为

(1δ)x1x222F(x1x2)22(1+δ)x1+x222 \begin{aligned} (1-\delta) \lVert \boldsymbol x_1-\boldsymbol x_2\rVert_2^2 & \leq \lVert \mathbf F(\boldsymbol x_1-\boldsymbol x_2)\rVert_2^2 \\ & \leq (1+\delta)\lVert \boldsymbol x_1+\boldsymbol x_2\rVert_2^2 \end{aligned}

这里δ(0,1)\delta\in(0,1),是一个小常数。RIP表明,从F\mathbf F的投影将距离保留在了由1δ1-\delta1+δ1+\delta所约束的两个信号之间。对于各种随机矩阵F\mathbf F和稀疏向量x\boldsymbol x,这个性质有很高的概率保留。这个性质保证在x\boldsymbol x稀疏的约束下将测量误差最小化,以很高的概率得到精确的重建x^x\hat{\boldsymbol x}\approx \boldsymbol x。重建公式为

x^=arg minxmFx22 \hat{\boldsymbol x} = \argmin_{\boldsymbol x} \lVert \boldsymbol m-\mathbf F\boldsymbol x \rVert_2^2

这个约束优化问题很难计算,但采样过程相对简单。

2.3.2 Compressed Sensing using Generative Models

稀疏性的要求限制了CS的应用,傅立叶基和小波基也只是部分缓解了这种限制,因为它们只能用于已知在这些基中稀疏的域。Bora et al., 2017提出使用生成模型的压缩感知(Compressed Sensing using Generative Models, CSGM)放宽这种要求。这个模型使用VAE或GAN预训练的深度神经网络GθG_\theta作为稀疏性的结构约束。这个生成器将一个潜在表示z\boldsymbol z映射到信号空间:x=Gθ(z)\boldsymbol x = G_\theta(\boldsymbol z)GθG_\theta 不要求稀疏信号,而是利用其架构和适应数据的权重将输出x\boldsymbol x隐式地约束在低维流形内。这种约束足以为随机矩阵提供广义的集限制特征值条件(Set-Restricted Eigenvalue Condition,S-REC),可以以较高的概率实现较低的重建误差。重建使用一个类似于CS的最小化过程:

z^=arg minzEθ(m,z)=arg minzmFGθ(z)22 \begin{aligned} \hat{\boldsymbol z} &= \argmin_{\boldsymbol z} E_\theta(\boldsymbol m,\boldsymbol z) \\ &= \argmin_{\boldsymbol z}\lVert \boldsymbol m - \mathbf FG_\theta(\boldsymbol z)\rVert_2^2 \end{aligned}

重建信号为x^=Gθ(z)\hat{\boldsymbol x} = G_\theta(\boldsymbol z)。上式arg min\argmin难以优化,可从随机采样点z^=pz(z)\hat{\boldsymbol z} = p_{\boldsymbol z}(\boldsymbol z)处开始进行梯度下降

z^=z^αEθ(m,z)zz=z^ \hat{\boldsymbol z} = \hat{\boldsymbol z} - \alpha\frac{\partial E_\theta(\boldsymbol m,\boldsymbol z)}{\partial \boldsymbol z}\bigg\lvert_{\boldsymbol z=\hat{\boldsymbol z}}

通常,需要数百或数千个梯度下降步骤以及从初始步骤重新启动几个步骤才能获得足够好的z^\hat{z}Bora et al., 2017; Bojanowski et al., 2018。这项工作建立了压缩感知与深度神经网络的联系,且优于LassoTibshirani, 1996Hand & Voroninski, 2017; , Dhar et al., 2018有在理论与实践上的进展。但仍有两方面的限制:

  1. 重建的优化需要上千步梯度下降,依然很慢
  2. 依赖于随机测量矩阵,对自然图像这样的高度结构化数据并不是最优的,经过学习的测量可能表现更好

2.3.3 Model-Agnostic Meta-Learning

元学习允许模型通过自我完善来适应新任务,与模型无关的元学习(MAML)提供了一种通用的方法来为许多任务调整参数Finn et al., 2017

文章提出的MAML可以用于任何使用梯度下降进行训练的问题和模型,除了深度神经网络,分类、回归、策略梯度强化学习等不同的结构和任务都可以在很小的改动下解决。在元学习中,训练模型的目标是从新的小规模样本中快速学习新任务,此外,模型由元学习器训练以有能力学习大量的不同任务。关键思路是训练模型的初始化参数,以使模型在新任务上通过计算少量新的数据进行一次或多次梯度步骤更新参数而拥有最好的性能。不像之前的思路,更新函数或学习规则,这个算法没有增加学习参数的数量,也不对模型做限制。

对每个任务:

θi=θαθLTi(fθ) \theta_i' = \theta - \alpha \nabla_\theta \mathcal{L_{T_i}}(f_\theta)

元学习:

θθβθTp(T)LTi(fθi) \theta \leftarrow \theta - \beta \nabla_\theta \sum_{\mathcal{T}\sim p(\mathcal{T})} \mathcal{L_{T_i}}(f_{\theta_i'})

2.3.4 Generative Adversarial Networks

生成对抗网络(GAN)训练参数化生成器GθG_\theta欺骗鉴别器DφD_\varphi,该鉴别器试图将真实数据与从生成器采样的虚假数据区分开来Goodfellow et al., 2014

minGmaxDV(D,G)=Expdata(x)[logD(x)]+Expdata(x)[log(1D(G(z))] \min_G \max_D V(D,G) = \mathbb{E}_{x \sim p_{data}(x)}[\log D(x)] + \mathbb{E}_{x \sim p_{data}(x)}[\log(1- D(G(z))]

2.3 深度压缩感知

首先,把 MAML 与 CSGM 相结合;之后将测量矩阵参数化,如使用 DNN。之前的工作使用随机映射作为测量,本文则将 RIP 作为训练目标,学习测量函数。之后通过在测量中施加 RIP 以外的其他性质,产生了两个新模型,包括带有判别器引导的潜在优化的 GAN 模型,导致更稳定的训练动态和更好的结果。

2.3.1 元学习压缩感知

CSGM 使用梯度下降,在隐空间内优化损失函数:

arg minzmFGθ(z)22 \argmin_{\boldsymbol z}\lVert \boldsymbol m - \mathbf FG_\theta(\boldsymbol z)\rVert_2^2

然而,这种优化可能需要上千个优化步骤,效率比较低。本文提出使用元学习的方法来优化这个重建过程。

3. Compressed Sensing using Generative Models

3.1 Abstract

The goal of compressed sensing is to estimate a vector from an underdetermined system of noisy linear measurements, by making use of prior knowledge on the structure of vectors in the relevant domain. For almost all results in this literature, the structure is represented by sparsity in a well-chosen basis. We show how to achieve guarantees similar to standard compressed sensing but without employing sparsity at all. Instead, we suppose that vectors lie near the range of a generative model G:RkRnG : \mathbb{R}^k \Rightarrow \mathbb{R}^n. Our main theorem is that, if G is L-Lipschitz, then roughly O(klogL)\mathcal{O}(k \log L) random Gaussian measurements suffice for an l1/l2l_1/l_2 recovery guarantee. We demonstrate our results using generative models from published variational autoencoder and generative adversarial networks. Our method can use 5-10x fewer measurements than Lasso for the same accuracy.

压缩感知的目标是通过利用相关域内向量结构的先验知识,从欠定的噪声线性测量系统中估计向量。在几乎所有相关文献的结果里,结构由在精心选择的基中的稀疏性所表示。我们展示了如何在完全不使用稀疏性的情况下实现与标准压缩感知相似的保证。与之相反,我们假设向量位于生成模型G:RkRnG : \mathbb{R}^k \Rightarrow \mathbb{R}^n的范围附近。我们主要的定理是,如果G是L-Lipschitz,那大约O(klogL)\mathcal{O}(k \log L)的随机高斯测量足以满足l1/l2l_1/l_2的恢复保证。我们使用已发布的变分自动编码器和生成对抗网络的生成模型展示了我们的结果。我们的方法在相同精度下比Lasso使用少5到10倍的测量。

3.2 Introduction

y=Ax+η y = Ax^* + \eta

其中,ARm×n,xRn,ηRmA\in \mathbb{R}^{m\times n}, x^*\in \mathbb{R}^n, \eta \in \mathbb{R}^m。这是一个欠定问题,求其最稀疏的解是一个NP难问题。如果A满足RIP或REC等条件,可以将其转化为凸优化问题,得到x的恢复。这方面有很多应用被研究,如计算断层扫描仪(CT),快速MRI等,以及单像素相机。

这篇文章依赖于生成模型产生的结构,而不是稀疏性。VAE、GAN等基于神经网络的生成模型在对数据分布建模上取得很多成果。这些模型的生成部件学习一个从低维表示空间zRkz\in\mathbb{R}^k到高维样本空间G(z)RnG(z)\in \mathbb{R}^n的映射。训练时,这个映射被鼓励产生与训练集相似的向量,因此可以使用预训练的生成器大致获取我们的域中“自然”向量的概念,因为生成器定义了样本空间向量上的一个分布。

主要创新:提出了一个使用生成模型进行压缩感知的算法。使用梯度下降优化表示zRkz\in \mathbb{R}^k,使得对应的图像G(z)G(z)有更小的测量误差AG(z)y22\lVert AG(z) - y\rVert_2^2。这个目标函数非凸,但我们发现梯度下降效果很好,结果在很少的测量下优于Lasso。

我们的理论结果表明,只要梯度下降可以找到我们目标的良好近似解,输出G(z)G(z)就可以几乎与真实的xx^*一样。

有定理:如果测量矩阵在给定生成器G的范围内满足S-REC,那么最小化测量误差的最佳值接近真实的xx^*。进一步地,随机高斯测量矩阵很大概率满足大部分的生成器的S-REC条件。

<!-- TODO -->

4. Task-Aware Compressed Sensing with Generative Adversarial Nets

4.1 Abstract

In recent years, neural network approaches have been widely adopted for machine learning tasks, with applications in computer vision. More recently, unsupervised generative models based on neural networks have been successfully applied to model data distributions via low-dimensional latent spaces. In this paper, we use Generative Adversarial Networks (GANs) to impose structure in compressed sensing problems, replacing the usual sparsity constraint. We propose to train the GANs in a task-aware fashion, specifically for reconstruction tasks. We also show that it is possible to train our model without using any (or much) non-compressed data. Finally, we show that the latent space of the GAN carries discriminative information and can further be regularized to generate input features for general inference tasks. We demonstrate the effectiveness of our method on a variety of reconstruction and classification problems.

近年来,神经网络方法已广泛用于机器学习任务,并应用于计算机视觉。最近,基于神经网络的无监督生成模型已成功应用于通过低维潜在空间进行数据分布建模。在本文中,我们使用生成对抗网络(GAN)在压缩感知问题中施加结构,从而取代了通常的稀疏约束。我们提出以任务感知方式训练GAN,尤其是针对重建任务。我们还表明,无需使用任何(或大量)非压缩数据即可训练模型。最后,我们证明了GAN的潜在空间带有判别信息,并且可以进一步进行正则化以生成用于一般推理任务的输入特征。我们证明了我们的方法在各种重构和分类问题上的有效性。

4.2 Introduction

这篇文章是在 [Bora et al.][1] 的基础上进行的工作。[Bora et al.] 使用生成模型替代了压缩感知问题中的稀疏性假设,并且证明在生成模型满足一定条件,且测量矩阵为独立同分布的随机高斯分布时,在很高的概率下有

G(z^)x26minzG(z)x2+3ζ2+2ϵ(1) \lVert G(\hat{z}) - x \rVert_2 \leq 6\min_z \lVert G(z) - x \rVert_2 + 3\lVert \zeta \rVert_2 + 2 \epsilon \tag{1}

[Bora et al.] 的文章中,损失函数取为

loss(z)=yAG(z)22+λpriorz22(2) loss(z) = \lVert y - AG(z) \rVert_2^2 + \lambda_{prior} \lVert z\rVert_2^2 \tag{2}

[Kabkab et al.] 首先证明了,在一定的假设下,有

limtEx[minzGt(z)x2]=0(3) \lim_{t\to \infty} \mathbb{E}_x \big[ \min_z\lVert G_t(z) - x \rVert_2 \big] = 0 \tag{3}

这表明 (1) 式右侧非常小。但以上定理的假设太严格,很难达到,因此 [Kabkab et al.] 考虑进行任务感知的 GAN 训练,允许生成器 G 对压缩感知重建任务进行优化。[Kabkab et al.] 对 Goodfellow 提出的的 GAN 的训练过程进行了修改,增加了对 (2) 式损失函数进行 GD 优化的步骤,再利用得到的隐变量 zz 继续 D 与 G 的优化。通过这种方式将压缩感知引入到 GAN 的训练中,使用压缩感知指导 GAN 的隐变量的筛选,从而得到更好的 GAN。

更进一步地,考虑到在压缩感知的任务中对采样过程进行压缩,非压缩的数据往往难以获得,因此提出基于一小部分的非压缩数据及大部分的压缩数据作为训练集进行训练。此算法训练两个判别器和一个生成器,第一个判别器分辨实际的训练数据 xx 与生成数据 G(z)G(z),第二个判别器分辨实际的压缩训练数据 yy 与生成数据 AG(z)AG(z)。这种对原始GAN的训练算法的变化并不会影响生成器的表示能力。

因为隐变量 zz 比非压缩数据 xx 与压缩数据 yy 的维度更低,作者认为可以将 zz 作为特征应用到分类任务中,并为这类任务引入了对比损失函数 (Contrastive Loss)。

作者在 MNIST、F-MNIST、CelebA 三个数据集上进行了实验。

5. Compressed Sensing with Invertible Generative Models and Dependent Noise

5.1 Abstract

We study image inverse problems with invertible generative priors, specifically normalizing flow models. Our formulation views the solution as the Maximum a Posteriori (MAP) estimate of the image given the measurements. Our general formulation allows for non-linear differentiable forward operators and noise distributions with long-range dependencies. We establish theoretical recovery guarantees for denoising and compressed sensing under our framework. We also empirically validate our method on various inverse problems including compressed sensing with quantized measurements and denoising with dependent noise patterns.

我们研究具有可逆生成先验的图像逆问题,特别是归一化流量模型。我们的公式将解决方案视为给定测量值的图像的最大后验(MAP)估计值。 我们的一般公式允许非线性可微正向算子和具有长期依赖性的噪声分布。我们在我们的框架下为消噪和压缩感知建立了理论上的恢复保证。我们还根据经验验证了我们的方法在各种反问题上的应用,包括采用量化测量值进行压缩感测以及采用相关噪声模式进行降噪的方法。

5.2 Introduction

5.3 预备知识

5.3.1 可逆生成模型

5.3.2 相关工作

5.4 方法

MAP 公式

与之前工作的联系

5.5 理论分析

不考虑信号的特定结构如稀疏与低维高斯生成先验。考虑普通设定,研究预训练流模型如何降低去噪问题的误差或减少压缩感知问题的测量数。

去噪问题的理论保证

压缩感知的理论保证

5.6 实验

(1)非线性正向算子,(2)具有相关性的复杂噪声,以及(3)高维可伸缩性。

数据集与结构

基线方法

模型平滑参数

结果

5.7 结论

6. Back-Projection based Fidelity Term for Ill-Posed Inverse Problems

不适定线性逆问题的基于反投影的保真项

6.1 Introduction

Ill-posed linear inverse problems appear in many image processing applications, such as deblurring, super- resolution and compressed sensing. Many restoration strategies involve minimizing a cost function, which is composed of fidelity and prior terms, balanced by a regularization parameter. While a vast amount of research has been focused on different prior models, the fidelity term is almost always chosen to be the least squares (LS) objective, that encourages fitting the linearly transformed optimization variable to the observations. In this paper, we examine a different fidelity term, which has been implicitly used by the recently proposed iterative denoising and backward projections (IDBP) framework. This term encourages agreement between the projection of the optimization variable onto the row space of the linear operator and the pseudo- inverse of the linear operator (“back-projection”) applied on the observations. We analytically examine the difference between the two fidelity terms for Tikhonov regularization and identify cases (such as a badly conditioned linear operator) where the new term has an advantage over the standard LS one. Moreover, we demonstrate empirically that the behavior of the two induced cost functions for sophisticated convex and non-convex priors, such as total-variation, BM3D, and deep generative models, correlates with the obtained theoretical analysis

不适定的线性逆问题出现在许多图像处理应用中,例如去模糊,超分辨率和压缩感知。许多恢复策略涉及使成本函数最小化,该成本函数由保真度和先验项组成,并由正则化参数进行平衡。大量研究集中在不同的先前模型上,但保真度术语几乎总是被选为最小二乘(LS)的目标,这鼓励将线性变换的优化变量拟合到观测值。在本文中,我们检查了一个不同的保真度项,该项已被最近提出的迭代去噪和反向投影(IDBP)框架隐式使用。该项鼓励优化变量在线性算子的行空间上的投影与应用于观测值的线性算子的伪逆(“反投影”)之间的一致。我们分析了两个保真项之间的差异,以进行Tikhonov正则化,并确定了新项比标准LS项具有优势的情况(例如条件不佳的线性算子)。此外,我们通过经验证明,复杂的凸和非凸先验的两个诱导成本函数的行为,例如总变异,BM3D和深度生成模型,与所得理论分析相关。