四、Basic Methods for Image Restoration and Identification（笔记&链接）

•What’s image degration?

Recorded image invariably represents a degraded version of the original scene. The undoing of these imperfections is crucial to many of the subsequent image processing tasks. There exists a wide range of different degradations that have to be taken into account, covering for instance noise, geometrical degradations, illumination and color imperfections and blur.

•What’s the reason for degration?

It can be caused by relative motion between the camera and the original scene, or by an optical system that is out of focus. When aerial photographs are produced for remote sensing purposes, blurs are introduced by atmospheric turbulence(大气湍流), aberrations(像差) in the optical system.

什么是图像退化？

录制的图像始终代表原始场景的降级版本。这些缺陷的消除对许多后续的图像处理任务至关重要。存在各种不同的降解，必须考虑到，覆盖例如噪音，几何降解，照明和颜色瑕疵和模糊。

退化的原因是什么？

它可能是由相机和原始场景之间的相对运动引起的，也可能是由不聚焦的光学系统引起的。当航空照片用于遥感时，大气湍流（大气湍流）、光学系统中的像差（像差）会引入模糊。

The field of image restoration (sometimes referred to as image deblurring or image deconvolution) is concerned with the reconstruction or estimation of the uncorrupted image from a blurred and noisy one.

Image restoration often uses some priori to establish a degrading system. Then it tries to perform an operation on the blurred image that is the inverse of the imperfections in the image degrading system.

图像恢复领域（有时称为图像去模糊或图像反卷积）涉及从模糊和噪声的图像重建或估计未损坏的图像。

图像恢复通常使用一些先验的方法来建立一个退化系统。然后尝试对模糊图像进行操作，这是图像退化系统中缺陷的反比。

Blur identification

In the use of image restoration methods, the characteristics of the degrading system and the noise are assumed to be known as priori. In practical situations, however, one may not be able to obtain this information directly from the image formation process.

The goal of blur identification is to estimate the attributes of the imperfect imaging system from the observed degraded image itself prior to the restoration process.

The combination of image restoration and blur identification is often referred to as blind  image deconvolution.

在图像恢复方法中，退化系统的特征和噪声被认为是先验的。然而，在实际情况下，人们可能无法直接从图像形成过程中获得这些信息。

模糊识别的目的是在图像恢复过程之前，从观测到的退化图像本身来估计不完全成像系统的属性。

图像恢复和模糊识别的结合通常被称为盲图像反卷积。

一、Blur Models

基本分为离散的点和连续的两种psf

Point-spread function（点扩散函数）

•d(x, y) takes on nonnegative values only, because of the physics of the underlying image formation process;

•when real-valued images are dealt with the point-spread function d(x, y) is real-valued too;

•the imperfections in the image formation process are modeled as passive operations on the data, i.e, no “energy” is absorbed or generated.

–2.1 NoBlur

PSF is modeled as a unit pulse:

–2.2 Linear Motion Blur

Many types of motion blur can be distinguished, all of which are due to relative motion between the recording device and the scene. This can be in the form of a translation, a rotation, a sudden change of scale, or some combinations of these. Here only the important case of a global translation will be considered.

Uniform motion in a straight line is the most general and common one of all the motion blurs. Others can be regarded as  piecewise Uniform motion in a straight line.

可以区分许多类型的运动模糊，所有这些都是由于记录设备和场景之间的相对运动造成的。这可以是平移、旋转、比例突然变化或这些组合的形式。这里只考虑全球翻译的重要情况。

直线匀速运动是所有运动中最普遍、最常见的运动模糊之一。其他的可以看作是直线上的分段匀速运动。

–2.3 Uniform Out-of-Focus Blur（均匀离焦模糊）

When a camera images a three-dimensional (3-D) scene onto a 2-D imaging plane, some parts of the scene are in focus while other parts are not. If the aperture(光圈) of the camera is circular, the image of any point source is a small disk, known as the circle of confusion (COC). However, if the degree of defocusing is large relative to the wavelengths considered, a geometrical approach can be followed resulting in a uniform intensity distribution within the COC, named uniform out-of –focus blur.

当相机在二维成像平面上对三维（3-D）场景进行成像时，场景的某些部分处于聚焦状态，而其他部分则不处于聚焦状态。如果相机的光圈是圆形的，则任何点光源的图像都是一个小圆盘，称为混淆圆（coc）。但是，如果离焦度相对于所考虑的波长较大，则可以采用几何方法，从而在COC内形成均匀的强度分布，称为均匀离焦模糊。

The diameter of defocus depends on :

1.D:diameter of lens

2.v0:image distance

3.f: the focal distance

4.u:the object distance

有半径R的这种均匀离焦模糊的空间连续PSF由下式给出：

同样对于这个PSF，离散版本d（n1，n2）不容易达到。 粗略近似是以下空间离散PSF：

其中C是必须选择的常数，以便d的积分等于1。

–2.4 Atmospheric Turbulence Blur（大气湍流模糊）

σG:determines the amount of spread of the blur

C:C  is  to  be  chosen  so  that  the integration（积分） of d is equal to 1.

二、Image Restoration Algorithms

小写f，h，g是离散模型的，大写F，H，G是连续模型的（H，h代表恢复图像但函数，和扩散函数有关）

–3.0 how to measure

Quantitative judgment:

   In image restoration the improvement in quality of the restored image over the recorded blurred one is measured by the signal-to-noise ratio(SNR) improvement.

https://www.cnblogs.com/qrlozte/p/5340216.html 信噪比

Visual judgment

   In image restoration the improvement in quality of the restored image over the recorded blurred one is measured by the observers’ eyes.

g是模糊图，f是理想图，f'是还原图

–3.1 Inverse Filter

•If the noise is absent, the second term above disappears so that the restored image is identical to the ideal image.

上面的第二项消失，因此恢复的图像与理想图像相同。

•The inverse filter may not exist because D(u,v) is zero at selected frequencies. This happens for both the linear motion blur and the out-of-focus blur described in the previous section.

逆滤波器可能不存在，因为D（u，v）在所选频率处为零。 对于上一节中描述的线性运动模糊和离焦模糊都会发生这种情况。

•Even if the blurring function’s spectral representation D(u,v) doesn’t actually go to zero but becomes small, the second term   -known as the inverse filtered noise-will become very large.

即使D存在但可能很小，这就导致第二项可能会很大。

–3.2 Least-Squares Filters

For the noise sensitivity of the inverse filter to be overcome, a number of restoration filters have been developed; these are collectively called least-squares filters. We describe the two most commonly used filters, namely the Wiener filter and the constrained least-squares filter.

为了克服逆滤波器的噪声灵敏度，已经开发了许多恢复滤波器; 这些统称为最小二乘滤波器。 我们描述了两种最常用的滤波器，即维纳滤波器和约束最小二乘滤波器。

The Wiener filter

The Wiener filter is a linear spatially invariant filter ,in which the point-spread function h(n1, n2) is chosen such that it minimizes the mean-squared error (MSE) between the ideal and the restored image.

维纳滤波器是线性空间不变滤波器，其中选择点扩展函数h（n1，n2）使得它最小化理想图像和恢复图像之间的均方误差（MSE）。

Here D* (u, v) is the complex conjugate(共轭复数) of D(u,v)。Sf(u, v) and Sw(u,v) are the power spectrum(功率谱) of the ideal image and the noise, respectively. The power spectrum is a measure for the average signal power per spatial frequency carried by the image.

Sf（u，v）和Sw（u，v）分别是理想图像和噪声的功率谱（功率谱）。 功率谱是图像携带的每空间频率的平均信号功率的量度。

Often the solutions to Sf(u, v) and Sw(u,v) are difficult to get, so we use a constant γ to replace      (信噪功率比).

In the noiseless case we have Sw(u,v) = 0, so that the Wiener filter approximates the inverse filter.

Constrained least-squares filter(正则滤波)

-The difference between the blurred version of restored image and the distorted one should equal to noisy data.

-Select the solution that is as “smooth” as possible

恢复图像的模糊版本与失真图像之间的差异应等于噪声数据。
选择尽可能“平滑”的解决方案

-The difference between the blurred version of restored image and the distorted one should equal to noisy data.

恢复图像的模糊版本与失真图像之间的差异应等于噪声数据。

Wiener filter tries to make the blurred version of the restored image equal to the recorded distorted image. But there exists noise. So constrained least-squares filter make this regular.

维纳滤镜试图使恢复图像的模糊版本等于记录的失真图像。 但是存在噪音。 所以约束最小二乘滤波器使这种规则成为常规。

Here α is a tuning or regularization parameter which makes the constraint satisfied. Though analytical approaches exist to estimate α, the regularization parameter is usually considered user tunable.

–3.3 Iterative Filters

迭代恢复滤波器的基本形式是迭代地接近逆滤波器的解。

If the number of iterations becomes very large, then fi(n1, n2) approaches the solution of the inverse filter:

三、Blur Identification Algorithms

In the previous section it was assumed that the point-spread function d(nl,n2) of the blur was known. In many practical cases the actual restoration process has to be preceded by the identification of this point-spread function. If the reasons for the blur are known, we could, in theory, determine the PSF analytically. Such situations are, however, rare. A more common situation is that the blur is estimated from the observed image itself. 

  The goal of blur identification is to estimate the attributes of the imperfect imaging system from the observed degraded image itself prior to the restoration process.

在前一部分中，假设模糊的点扩散函数d（n1，n2）是已知的。 在许多实际情况下，实际的恢复过程必须先识别这种点扩散函数。 如果已知模糊的原因，我们理论上可以分析地确定PSF。 然而，这种情况很少见。 更常见的情况是从观察到的图像本身估计模糊。
   模糊识别的目标是在恢复过程之前从观察到的劣化图像本身估计不完美成像系统的属性。

–4.1 Spectral Blur Estimation

We have seen that two important classes of blurs, namely motion and out-of-focus blur, have spectral zeros. Since the degraded image is described by the equal as follows,

The spectral zeros of the PSF should also be visible in the Fourier transform G(u,v). The structure and location of the zero patterns can be estimated. The structure of the zero patterns characterizes the type and degree of blur.

In the case in which the pattern contains dominant parallel lines of zeros, an estimate of the length and angle of motion can be made. In case dominant circular patterns occur, out-of-focus blur can be inferred and the degree of out of focus can be estimated.

在图案包含主要的零平行线的情况下，可以估计长度和运动角度。 在出现主要圆形图案的情况下，可以推断出离焦模糊并且可以估计失焦的程度。

–4.2 Maximum-Likelihood Blur Estimation

In case the point-spread function does not have characteristic spectral zeros or in case a parametric blur model such as motion or out-of-focus blur cannot be assumed, the individual coefficients of the point-spread function have to be estimated.

Most maximum-likelihood identification techniques begin by assuming that the ideal image can be described with the 2-D autoregressive(自回归) model as follows,

The parameters of this image model—that is, the prediction coefficients     and the variance   of the white noise      —are not necessarily assumed to be known.