问题描述:

相机晃动造成的运动模糊是摄影中一个普遍存在的操作问题。在很多场合中,譬如取景处光线太暗而不得不增加快门时间,或者拍摄的物体运动太快,都不可避免地造成照片的模糊,影响最终效果。Deblurring技术就是将模糊的照片恢复成更清晰或分辨率更高的照片。

One of the most common artifacts in digital photography is motion blur caused by camera shake. In many situations,  there simply is not enough light to avoid using a long shutter speed, or  the motion of objects is too fast to capture, and  the inevitable result is that many of our snapshots come out blurry and disappoint-ing.Deblurring is a technique to recover an un-blurred image from a blurred photogragh.

问题分类:

单一去模糊:从单一图片去除模糊

Single-image Deblurring:

多图去模糊:拥有同一场景多张模糊图像,使得模糊点更加可靠和易于处理 

Multi-image Deblurring:

 

Deblurring算法分类:

  • deconv method

If a motion blur is shift-invariant, it can be modeled as the convolution of a latent image with a motion blur kernel, where the kernel describes the trace of a sensor. Then, removing a motion blur from an image becomes a deconvolution operation.

如果一个运动模糊是平移-不变的,它可以被建模为一个带有运动模糊内核的潜在图像的卷积,在那里,“科尔”描述了一个传感器的轨迹。然后,从图像中移除一个运动模糊就变成了解卷积操作。

Delurring 综述-LMLPHP

Image deconvolution can be further separated into the blind and non-blind cases.

图像解卷积 可进一步分类为盲和非盲两种方法。

In non-blind deconvolution, the motion blur kernel is assumed to be known or computed elsewhere.

在非盲解卷积中,运动模糊核是已知的。

the only task remaining is to estimate the unblurred latent image.

剩下的唯一任务是估计出不模糊的潜像。

Blurred image=Sharp image* kernel+Noise  
Delurring 综述-LMLPHP

 

In blind deconvolution, the kernel is unknown and the recovery of the latent image becomes more challenging.

盲解卷积,由于内核未知,潜像的恢复更具挑战性。

Single-image blind deconvolution is an ill-posed problem because the number of unknowns exceeds the number of observed data.

单图像盲解卷积是一个非常困难的问题,因为未知数据远远多于可观测的数据。

Most blind deconvolution methods take an iterative process that alternatingly optimizes the motion blur kernel and the latent image.In the process, the blur kernel is obtained from the estimated latent image and the given blurred image. The kernel is then used to estimate the latent image by applying non-blind deconvolution to the given blurred image. The new estimated latent image is used for kernel estimation in the next iteration. The intensive computation of previous methods stems from the complicated methods used for kernel estimation and latent image estimation. Optimization involving large matrices and vectors is needed for kernel estimation, and sophisticated optimization techniques are necessary to handle non-blind deconvolution with non-linear priors.

大多数盲解卷积方法采用的是迭代地交替优化运动模糊核和潜像。在这个过程中,模糊核由估计的潜像以及给出的模糊图像算得。之后这个模糊核用于非盲解卷积来估计模糊图像的潜像。新估计的潜像将用于下次迭代的核估计。由于核估计和潜像估计过程的复杂性,以前的算法都需要密集的运算力。核估计需要对大型矩阵和向量做优化,若想解决非线性先验的非盲解卷积过程,精妙的优化技巧也是必要的。

 

  • base gyro method
  • feature method

 

相关工作:

Single-image blind deconvolution 相关:

1996 《Image motion estimation from motion smear - a new computa-

tional model》

1998《Total Variation Blind Deconvolution: The Devil is in the Details》

2007《Removing non-uniform motion blur from images》

2008《High-quality Motion Deblurring from a Single Image》

 

 

10-07 13:30