大部分的單一影像盲去模糊演算法是使用整張模糊影像及粗略到精細的迭代步驟去估計並找出最大化後驗機率的解，也就是模糊核心(又叫點擴散函數)，我們觀察到利用整張影像去估計出模糊核心通常需要大量又複雜的計算，在迭代步驟運算上會花費許多時間；而這種方法得出的清晰影像不一定都能有最好的效果，影像中不重要的部份可能會導致估計的模糊核有錯誤。
為了避免這些問題，在這篇論文中，我們提出一個利用影像分解及區塊選取方法選出對估計模糊核心有用的區塊。首先，我們計算出模糊影像的梯度值並利用自然影像的梯度值分佈資訊來對模糊影像做分解，將影像區分出包含有用資訊的結構部分及沒有幫助的平滑區域。接著從結構部分找出具有強邊緣的有用資訊像素(若影像中含有飽和區域，我們也會另外找出飽和區的有用資訊像素點)，將距離相近的資訊點各自形成不同大小的區塊，再利用區塊的大小及區塊內資訊點的像素值等條件去計算每個區塊的效用值，最後選出擁有最大效用值的區塊作為估計模糊核心的初始區塊。我們的實驗結果顯示我們的方法在真實的模糊影像及合成的模糊影像上都能節省去模糊的時間並保有不錯的去模糊品質。

Most of single image blind deblurring algorithms use the whole image to search for the maximum a posteriori (MAP) solution to estimate the blur kernel for the blurred image based on the coarse-to-fine iterations. We observed that using the whole blurred image to estimate its blur kernel usually needs a lot of complicated calculation. The quality of the latent image obtained from this method is not always good because the small gradients of the image may lead to a lot of errors in the kernel estimation.
We propose an algorithm based on image decomposition and patch selection to select the helpful patch for the kernel estimation to avoid the problems given above. First, we use the gradient magnitudes of the blurry photo and the distribution of gradients within a natural image to decompose the blurry image into informative structure and insignificant smooth region. Second, we find the most informative pixels from the structure region. If the image contains saturation regions, we will also find the useful pixels within these regions. These pixels will form different sizes of patches with their nearest neighbor. We use the size of patch and the values of informative pixels within this patch to calculate the utility of each patch. Finally, we select the patch with the maximum utility as the initial patch for the kernel estimation. Our experimental results show that our algorithm can save time for the deblurring while preserving the good qualities of deblurred photos both on real-world and synthetic images.

[1] Fergus, R., Singh, B., Hertzmann, A., Roweis, S.T., Freeman, W.T. Removing camera shake from a single photograph. ACM Trans. Graph. Vol. 25, No. 3, pp. 787–794 (2006)
[2] Shan, Q., Jia, J., Agarwala, A. High-quality motion deblurring from a single image. ACM Trans. Graph. Vol. 27 (2008)
[3] Cai, J.F., Ji, H., Liu, C., Shen, Z. Blind motion deblurring from a single image using sparse approximation. In CVPR, pp. 104–111 (2009)
[4] Yuan, L., Sun, J., Quan, L., Shum, H.Y. Image deblurring with blurred/noisy image pairs. ACM Trans. Graph. Vol. 26, No. 1 (2007)
[5] Yuquan Xu, Xiyuan Hu, Lu Wang, Silong Peng. Single image blind deblurring with image decomposition. ICASSP, IEEE Conference on pp. 929-932 (2012)
[6] Xu, L., Jia, J. Two-phase kernel estimation for robust motion deblurring. In Computer Vision–ECCV 2010. pp. 157-170. Springer Berlin Heidelberg. (2010)
[7] Jong-Ho Lee , Yo-Sung. High-quality non-blind image deconvolution. 4th PSIVT, IEEE Conference on pp.282-287 (2010)
[8] Yoav HaCohen, Eli Shechtman, Dani Lishchinski. Deblurring by example using dense correspondence. ICCV, IEEE Conference on pp. 2384-2391 (2013)
[9] Joshi, N., Szeliski, R., Kriegman, D.J. Psf estimation using sharp edge prediction. In CVPR (2008)
[10] Cho, S., Lee, S. Fast motion deblurring. ACM Trans. Graph. Vol. 28, No. 5,p.145 (2009)
[11] Köhler, R., Hirsch, M., Mohler, B., Schölkopf, B., Harmeling, S. Recording and playback of camera shake: Benchmarking blind deconvolution with a real-world database. In Computer Vision–ECCV 2012. pp. 27-40. Springer Berlin Heidelberg. (2012)
[12] Levin, A., Weiss, Y., Durand, F., Freeman, W. T.. Understanding and evaluating blind deconvolution algorithms. In CVPR, 2009, IEEE Conference on pp. 1964-1971 (2009)