李明穗臺灣大學:資訊工程學研究所謝若元Hsieh, Jo-YuanJo-YuanHsieh2010-06-092018-07-052010-06-092018-07-052009U0001-1808200911294800http://ntur.lib.ntu.edu.tw//handle/246246/185432在這篇論文中，我們針對兩張影像來探討，一張模糊以及一張清楚且經過幾何轉換的影像。我們試著拼接兩張影像並試著去模糊化。長久以來，由於去模糊是個在不適定的情況下求逆轉換的問題，所以對影像去模糊一直是個極具挑戰性的任務。近年來不論是利用多張或是單張影像的去模糊法都被廣泛的提出討論，其中這些方法又可被分為兩類：盲去模糊法及非盲去模糊法。如果模糊核在去模糊的過程中為未知，則被稱為盲去模糊；反之，若模糊核在去模糊過程中被假設為已知前提，則稱之為非盲去模糊。為了算出模糊核，我們打算利用兩張影像中清楚的那張來幫助計算。理論上，直接將這兩張影像接在一起便可得到重疊的兩塊區域，一塊為模糊而一塊為清楚。藉由這重疊且被對在一起的兩塊區域，我們能夠算出影像的模糊核。但是實際上由於模糊的影像有些資訊已經被破壞，直接將兩張影像接在一起無法得到準確的重疊，因此我們決定先將模糊影像在拼接前先經過一次去模糊，將去模糊後的影像與清楚影像相接，並記其錄轉換參數，再將模糊影像和清楚影像以方才紀錄的參數相接，便可得到準確的重疊區塊。也因此模糊核能夠被準確的計算出來。最後便可以利用被計算出來的模糊核以非盲去模糊法來回復。In this thesis, we try to stitch one clear image with a blurred image relative to a geometric transformation, and then recover the blurred image in the meantime. Image deblurring has long been a challenging work since it is an ill-posed inverse problem. Deblurring methods using multiple or single image are both discussed in recent years. The deblurring is called blind if the kernel is unknown or non-blind if the kernel is known a priori. In order to estimate the blur kernel, we try to take the information from the non-blurred patch for help. By stitching a blurred image with a non-blurred image using Speeded-Up Robust Features (SURF), we can obtain the aligned overlapped patches. Ideally, we can estimate the blur kernel based on blurred/non-blurred patches. However, directly stitching blurred/non-blurred images leads to poor aligned patches. As a result, the kernel is misestimated and the image is incorrectly recovered. To solve this issue, a pre-deblurring as a pre-processing step of the blurred image is considered. We stitch the pre-deblurred image with the non-blurred image and record the transformation parameters for temporary. After that we stitch the original blurred image with the non-blurred image using the recorded parameters to get better-aligned patches. Now the two patches are much better-aligned than before so that the kernel can be correctly estimated. Finally, promising result using progressive inter-scale and intra-scale deconvolution is presented.口試委員會審定書謝 i文摘要 iiBSTRACT iiiONTENTS ivIST OF FIGURES vihapter 1 Introduction 1.1 Motivation 1.2 Problem Statement 2.3 Thesis Organization 3hapter 2 Related Work 4.1 Interest Points Detection 4.1.1 Moravec Corner Detection 4.1.2 Harris Corner Detection 5.1.3 Scale Invariant Feature Transform (SIFT) 6.2 Image Deblurring 7.2.1 Blind Image Deconvolution 7.2.2 Non-blind Image Deconvolution 8hapter 3 Background Knowledge 10.1 Speeded-Up Robust Feature (SURF) 10.1.1 Interest Point Detection 11.1.2 Interest Point Description 14.1.3 Matching 16.2 Weighted Least Square Estimation of Transformation Parameters Using Expectation Maximization Algorithm 17.2.1 Least Square Estimation of Transformation Parameters 17.2.2 Weighted Least Square Estimation using EM Algorithm 19.3 Kernel Estimation 20.3.1 Kernel Estimation using Tikhonov Regularization 20.4 Image Deconvolution 21.4.1 Richardson-Lucy (RL) Algorithm 21.4.2 Bilateral Richardson-Lucy (BRL) Algorithm 22.4.3 Progressive Inter-scale Scheme 23.4.4 Joint Bilateral Richardson-Lucy (JBRL) Algorithm 24.4.5 Progressive Intra-scale Scheme 24hapter 4 Proposed System 27.1 System Overview 27.2 Naive approach (case1) 29.3 Pre-deblur approach (case2) 33hapter 5 Experiment and Results 37.1 Deblur with Directly Stitching Blurred/Non-blurred images 37.2 Deblur with Stitching Pre-deblurred/Non-blurred images 45hapter 6 Conclusion and Discussion 58.1 Conclusion 58.2 Discussion 59EFERENCE 657050146 bytesapplication/pdfen-US影像去模糊影像拼接全景畫Image deblurringimage stitchingpanoramathesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/185432/1/ntu-98-R96922043-1.pdf