高解析影像與光場影像之去模糊技術

Image Deblurring Technologies for Large Images and Light Field Images

王能謙(Neng-Chien Wang)

指導教授 ： 丁建均

國立臺灣大學/電機資訊學院/電信工程學研究所/碩士(2016年)

https://doi.org/10.6342/NTU201610184

摘要

影像處理是個熱門的研究課題，本篇論文分為兩個研究方向，首先介紹的是模糊影像方面的技術，次為光場相機的應用技術。
目前已有許多影像解模糊的文獻，其中研究可分為二：盲目解迴旋、非盲目解迴旋，這篇主要探討為在高解析影像下有效率的解決非盲目解迴旋的問題。
高解析影像的資料量使得解模糊運算複雜度大量的提升，我們希望能以更短的時間來達成解模糊，所以採用了2009年Krishnan所提出的快速影像反摺積。為了解決高運算複雜度的問題，我們使用了區塊影像解模糊的方法，並透過運算複雜度找出最佳的子影像切割數量，再將子影像拼湊成完整的影像，但是直接拼湊會造成塊狀效應，所以子影像間必須部分重疊且給予線性權重，才能解決塊狀效應。而重疊部分的多寡決定了運算時間及成果，重疊部分越少運算時間越少，但成果越差。為了取得一個平衡，我們選擇了一個特定的重疊大小，能兼顧效率及成果。
另一個要介紹的是光場相機，光場相機單次拍照就能捕捉到一般影像並且額外提供影像的角度及位置資訊，藉此我們能重建場景深度，以此估測出影像內物體的景深。
光場相機是以鏡片陣列組成而產生子影像陣列，我們必須先計算每個子影像的偏移量，才能重建完整影像。其中我們會使用四元樹更精確計算出偏移量，白場影像來補償子影像周圍較暗的問題，使用影像繪圖法增加重建影像品質。
影像重建完成後，我們將以影像分割技術，對影像中的物體個別分割，再以每個物體為單位，藉由其偏移量估測出其景深，以此重建出整個影像的景深圖及每個物件的景深距離。

關鍵字

解模糊 ； 高解析影像 ； 光場 ； 四元樹 ； 影像繪圖法

並列摘要

Image processing has been developed for a long time. This paper can be separated into two parts. We will introduce the proposed techniques of image deblurring at first. Then the proposed light field deblurring algorithm will be introduced.
The literatures of image deblurring can be categorized into two classes: blind deconvolution and non-blind deconvolution. First, we try to improve the efficiency of non-blind deconvolution in ultra-high resolution images.
The complexity of deblurring is raised in ultra-high resolution images. Therefore, we try to reduce the computation time. We modified the algorithm “Fast Image Deconvolution” proposed by Krishnan in 2009. To reduce complexity, we process the image in block, and find the optimal division that can minimize the complexity. Merging the result of each block directly will cause blocking effect, so it should be overlapped between sub-images with linear weight. The size of overlapping decided our computing time and performance. Less overlapping is more efficient but leads to worse performance. For balance, we choose a specific size of overlapping which give consideration both efficiency and performance.
Another topic is light field deblurring. A light field camera can capture the location and the angle information from a single shot. Thereby, we can reconstruct the depth of scene and stereoscopic images can be obtained.
A light field camera is composed by the array of lens. We will obtain sub-images by every lens. If we want to render the image, we have to obtain the disparity of each microimage pair and hence we can estimate the information of the depth.
At first, we obtain the relationship among microlenses by using regression analysis. Then, we take white image into consideration to compensate the luminance of the edge of every microimage and use quad-trees to compute disparity more precisely. Moreover, we use the image-based rendering technique to improve the quality of the reconstructed image.
After rendering image, we use the technique of image segmentation. Then, every object will be cut apart. We estimate the depth of every object by the disparity, and hence we can reconstruct the depth map of the whole image.

並列關鍵字

deblurring ； ultra-high resolution image ； light field ； quad-tree ； image-based rendering

參考文獻

[1]Asif, Muhammad, Aamir Saeed Malik, and Tae-Sun Choi. "3D shape recovery from image defocus using wavelet analysis." Image Processing, 2005. ICIP 2005. IEEE International Conference on. Vol. 1. IEEE, 2005.

[4]Chen, Po-Chang, Yung-Lin Chen, and Hsin-Yueh Sung. "Image restoration based on multiple PSF information with applications to phase-coded imaging system." SPIE Optical Engineering+ Applications. International Society for Optics and Photonics, 2010.

[5]Richardson, William Hadley. "Bayesian-Based Iterative Method of Image Restoration*." JOSA 62.1, 55-59, 1972.

[9]Levin, Anat, et al. "Image and depth from a conventional camera with a coded aperture." ACM Transactions on Graphics (TOG). Vol. 26. No. 3. ACM, 2007.

[10]Shan, Qi, Jiaya Jia, and Aseem Agarwala. "High-quality motion deblurring from a single image." ACM Transactions on Graphics (TOG). Vol. 27. No. 3. ACM, 2008.