基於結構化森林與L0平滑的 影像去模糊

Image Deblurring via Structured Forest and L0 Gradient Minimization

潘科維(Ko-Wei Pan)

指導教授 ： 張隆紋

國立清華大學/電機資訊學院/資訊工程學系所/碩士(2018年)

摘要

利用單一模糊影像去模糊已被長時間討論，現今多數去模糊演算法都是使用整張影像的資訊來估算模糊核(kernel)，並利用最大後驗機率估計(Maximum the a posteriori, MAP)方式，將模糊影像由粗略到精細，經多次迭代，交替估測模糊核與隱式影像(latent image)直到收斂。
然而，利用整張影像來估算模糊核並不是永遠都能得到好的結果，影像中許多區域並不能提供正確的模糊資訊，有時反而會使估計結果有錯誤並耗費更多時間計算。因此，我們提出了一個基於邊緣資訊的區塊選擇方法，來選取對於估測模糊核有幫助的區塊，減少處理資訊量。
我們一開始將模糊影像和其對應清晰影像的邊緣進行分析，引用並改良 [1]，訓練出一個結構化森林(structured forest)，接著只要將模糊影像輸入訓練好的模組，就可以很快的找出多個含有豐富邊緣資訊的區塊，再利用空間資訊，選擇是否要合併，最後產生一個或多個對於估測模糊核較有幫助的區塊，再利用最大後驗機率估計模糊核。
在交替估測的時候，我們發現隱式影像中如果含有過多細節，亦可能會影響模糊核估測的正確性，所以在每次迭代前，先對影像取L0 平滑(L0 Gradient Minimization)，去掉多餘的細節並保留主要結構，再從結構中提取邊緣資訊進行模糊核估測。

關鍵字

影像去模糊 ； 結構化森林 ； 物件估測 ； L0梯度最小化

並列摘要

Recovering an un-blurred image from a single, motion-blurred photograph has long been a fundamental research problem in digital imaging. Recently, most deblurring algorithm use whole image as input to estimate the blur kernel based on a coarse-to-fine MAP approach.
However, not all pixels of the input blurred image are informative. Many regions in the images, smooth regions for example, do not contribute much for estimating the blur kernel, even take more time to compute. Thus, we present the regions selection method based on strong edges to automatically select the good regions for estimating the blur kernel.
At the beginning, we analyze the blurred images and the edges correspond with its clear images, retraining the structured forest based on [1]. The trained module takes the blurred image as input and find the candidate regions which contain strong edges. Then we use the spatial information to decide whether to merge these regions or not. After that, we can get the patch which can contribute much for estimating the blur kernel.
At the estimation time, we find that too much detail in the latent image may affect the correctness of the kernel. Consequently, we apply L0 Gradient Minimization on the latent image before each iteration to remove extra details and preserve main structure simultaneously. Then extract the edges from the structure to improve the quality of kernel estimation.

並列關鍵字

Image Deblurring ； Structure Forest ； Object Proposal ； L0 Gradient Minization

參考文獻

[1] P. Dollár and L. C. Zitnick, "Fast Edge Detection Using Structured Forests," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 37, no. 8, pp. 1558-1570, 1 Aug 2015.

[2] R. Fergus, B. Singh, A. Hertzmann, S. T. Roweis and W. T. Freeman, "Removing camera shake from a single photograph," Proceedings of ACM SIGGRAPH, vol. 25, no. 3, pp. 787-794, July 2006.

[3] A. Levin, Y. Weiss, F. Durand and W. T. Freeman, "Understanding and evaluating blind deconvolution algorithms," Proceedings of IEEE conference on computer vision and pattern recognition, pp. 1964-1971, June 2009.

[4] N. Joshi, R. Szeliski and D. J. Kriegman, "PSF estimation using sharp edge prediction," Proceedings of IEEE conference on computer vision and pattern recognition, June 2008.

[5] T. S. Cho, S. Paris, B. K. P. Horn and W. T. Freeman, "Blur kernel estimation using the radon transform," Proceedings of IEEE conference on computer vision and pattern recognition, pp. 241-248, June 2011.