去模糊在電腦視覺的領域上是一個研究已久的課題，需要扎實的理論基礎以及應用的實作性，通常可以將模糊的問題建構成摺積(convolution)的形式，而其中最具挑戰的就是藉由在單一張影像上的盲式反摺積(blind deconvolution)。許多文獻透過迴圈的技巧，交替的估測隱式影像(latent image)與模糊核(kernel)以達到估測模糊核，再運用非盲式反摺積(non-blind deconvolution)，還原出清晰影像。
本篇論文針對模糊影像，採用兩階段的方法估測出準確的模糊核。第一階段基於自然影像中L_0 norm的先驗(prior)條件，同時增加了對於區域平滑的限制，使用簡單的高斯濾波器(Gaussian filter)維持影像中平坦的地方；第二階段將估測模糊核進行校正，利用L_0 norm的特性，使得模糊核更加稀疏，藉此去除掉低強度(low intensity)像素的地方。
實驗結果顯示只要調整適當的參數，在不需要額外資料庫的條件下，對於還原出清晰人臉也可以有很好的效果。

Deblurring form a single blurred image is a challenge task in computer vision. It is an ill-posed problem to estimate the unknown blur kernel and recover the original image. There are many significant deblurring methods toward the natural images; however, few of them are not able to perform well on face images. Based on L_0 norm prior, we propose a two-step method for the images deblurring. The proposed method does not require any facial dataset to initialize the gradient of contours or any complex filtering strategies. In first step, we combine L_0 norm prior with our local smooth prior to predict the blur kernel. With simple Gaussian filtering, we could maintain the smooth region in the sharp image. In second step, refine the previous kernel result. In order to discard low intensity pixels (seemed to be noises) on kernel, we impose the sparsity on the kernel with L_0 norm regularization. Experimental results demonstrate that our proposed algorithm perform well on the facial images.

[1] R. Fergus, B. Singh, A. Hertzmann, S. T. Roweis, and W. T. Freeman. Removing camera shake from a single photograph. ACM Trans. Graph., 25(3):787–794, 2006.
[2] Q. Shan, J. Jia, and A. Agarwala. High-quality motion deblurring from a single image. ACM Trans. Graph., 27(3):73, 2008.
[3] S. Cho and S. Lee. Fast motion deblurring. SIGGRAPH ASIA, 2009.
[4] A. Levin, Y. Weiss, F. Durand, and W. T. Freeman. Understanding blind de-convolution algorithms. In CVPR, 2009
[5] L. Xu and J. Jia. Two-phase kernel estimation for robust motion deblurring. In ECCV, 2010 (pages 157–170).
[6] D. Krishnan, T. Tay, and R. Fergus. Blind deconvolution using a normalized sparsity measure. In CVPR, 2011 IEEE Conference on (pp. 233-240).
[7] A. Levin, Y. Weiss, F. Durand, and W. T. Freeman. Efficient marginal likelihood optimization in blind deconvolution. In CVPR, 2011
[8] H. Bae, C. C. Fowlkes, P. H. Chou. Patch mosaic for fast motion deblurring. In ACCV, 2012 (pp. 322-335).
[9] L. Xu, S. Zheng, and J. Jia. Unnatural l0 sparse representation for natural image deblurring. In CVPR, 2013 (pages 1107–1114).
[10] L. Sun, S. Cho, J. Wang, and J. Hays. Edge-based blur kernel estimation using patch priors. In ICCP, 2013.
[11] J. Kotera, F. Šroubek, and P. Milanfar. Blind deconvolution using alternating maximum a posteriori estimation with heavy-tailed priors. R. Wilson, E. Hancock, A. Bors, W. Smith, Eds., Springer: Berlin/Heidelberg, Germany, Volume 8048, In CAIP, 2013 (pages 59–66).
[12] T. Michaeli and M. Irani. Blind Deblurring Using Internal Patch Recurrence. In ECCV, 2014
[13] J. Pan, Z. Hu, Z. Su, and M.-H. Yang. Deblurring Text Images via L_0-Regularized Intensity and Gradient Prior. In CVPR, 2014 (pages 2901-2908).
[14] J. Pan, Z. Hu, Z. Su, and M.-H. Yang. Deblurring face images with exemplars. In ECCV, 2014.