利用深度差進行單張雙深度影像之去模糊化

Abstract

在用隨身相機拍攝有明顯深度差異的場景時，拍攝時的晃動會導致相片產生不均勻的模糊：與相機距離較近的物品會有較嚴重的模糊，反之與相機距離較遠的物品則是較為清晰。這篇論文提出了在雙深度的場景下來對單張模糊影像去模糊化的方法。我們有兩個假設：一是相機的晃動主要是在2D平面上的直線移動，在這個假設下，我們可以將模糊程度的不同單純歸因為場景深度的差異，而不用考慮轉動等其它因素所導致的不均勻模糊；二是相機拍攝的場景大致上要可以分成深度差很大的前景和背景，如此一來我們就可以根據模糊程度的不同來切割出前景和背景，分別去模糊化。我們的方法大致如下所述：先將影像切割成數個區塊，估計出每個區塊的blur kernel，接下來計算這些blur kernel與reference kernel的相似程度，據此將這些blur kernel分成兩群並分別估計出前景和背景的blur kernel。再來我們根據模糊程度的不同，將影像切割為前景和背景。最後再用疊代加權最小平方法(Iterative Re-weighted Least Square)解能量方程式，分別對前景和背景進行去模糊化，以還原出一張清晰的影像。

Camera motion on scenes with depth variation would cause spatial varying blur on image: the distance from the camera is smaller, the relative movement of projected image pixel is larger. The proposed method aims to deblur a single blurred image caused by large depth variation in the scene. We have two assumptions for our method: one is the camera motion is 2D translational motions, and then we do not need to consider spatially-variant motion blur caused by rotational motions or other camera motions; another is the blurred image can be roughly composed of foreground and background regions with large depth variation, which can make the kernel clustering and image segmentation much easily. In this paper, we propose a method to segment blurred image taken from scenes with large depth variation into foreground and background regions, and then estimate the representative blur kernel for each region individually. Finally, we use IRLS(Iterative Re-weighted Least Square) method to solve the energy function for reconstructing a clear image.