請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/55330
標題: | 應用於疏光場去除疊影事後對焦之演算法與架構設計 Anti-aliased Refocus Algorithm and Architecture for Sparse Light Field |
作者: | Ruei Jin 金睿 |
指導教授: | 陳良基 |
關鍵字: | 光場,事後對焦,影像生成,透鏡陣列相機,疊影效應,資料重複使用架構, light field,refocus,view synthesis,aliasing effect,data reuse scheme, |
出版年 : | 2015 |
學位: | 碩士 |
摘要: | 隨著科技的日新月異,突破傳統相機限制的光場相機在學術圈、科技界和攝影界掀起一波革命,而各種光場應用中,以事後對焦這項應用最廣為人知。相較於傳統平面影像,光場相機能提供較高品質的深度資訊和數位事後對焦的功能。即使近年來計算攝影學演算法的迅速發展,目前事後對焦的演算法仍然無法符合即時影像生成的需求。產生即時且高品質的事後對焦相片已成為現今研究中的一項挑戰。
本論文首要探討疏光場組態下會遇到的疊影問題,首先探討造成疊影效應的原因和各種解決方法,我們針對內插影像的方式提出改良,設計出一套二維線性掃描的演算法,可以用比現有方法快約二十倍的速度完成計算,因做了簡單化的假設使得品質有部分犧牲,但透過比較可以發現結果是肉眼上可以接受的品質犧牲,另外我們討論不同的場景和光場像機組態下,使用二維線性掃描演算法的計算表現和二維線性掃描演算法的運算量須提升到何種程度才能夠消除疊影。 接著討論在硬體實作上遭遇的問題,由於圖片解析度和相機陣列間的間距提高,視差值也隨之必須提高,造成為了解決疊影效應所需生成的內插影像就越多,而大部分的影像生成演算法因為訊息之間的相依性和補洞的程序,多數的影像生成演算法應用在事後對焦幾乎無法在硬體實作上達到實時的速度需求。二維線性掃描演算法能夠基於塊狀分割來達到減少靜態隨機存取記憶體的面積,另外我們分析了使用等級C資料重複使用架構來減少頻寬損耗,並分析不同的塊狀大小下資料重複使用架構的表現。 本論文並針對複雜度提出了二維線性掃描的快速計算方式,能在犧牲極少品質的狀況下,大量降低運算複雜度,保有效能,並能在硬體架構上實行。另外,提出硬體架構並實做了資料預讀的處理模組,節省了極大的頻寬損耗。 In recent years, a new research topic, computational photography refers to computational image capture, processing, and manipulation techniques that enhance or extend the capabilities of digital photography. Good examples of image capture and manipulation include panoramas, and high-dynamic-range imaging which is the use of differently exposed pictures of the same scene to extend dynamic range beyond even that of analog film-based media. Computational photography is a section between computer graphic and photography. Light field camera is also a hot topic in computational photography. The camera capture the 'lightrays' instead of pixels. The cameras could be great depth sensor, and can do a lot of novel applications which traditional cameras cannot do, such as refocus, 3D model reconstruct. However, digital refocusing for sparsely sampled light elds results in aliasing effect which will get worse for higher image resolution or larger baseline. Previous works perform anti-aliasing by view interpolation but mostly apply time-consuming algorithms. In this thesis we study this aliasing problem and then a fast refocus algorithm with simple view interpolation is presented. The algorithm we proposed is called 1D-line-scan and 2D-line-scan algorithm. Our 1D-line-scan can be about 20 times faster than VSRS-1D-fast. Finally, a novel block-based 2D-line-scan is proposed. We develop the block-based version is for being hardware-oriented and found it can map to the Level C data reuse scheme. We further analyse and implement a pre-loading module to reduce the bandwidth cost. Totally speaking, a highly effective algorithm for digital refocusing is designed. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/55330 |
全文授權: | 有償授權 |
顯示於系所單位: | 電子工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-104-1.pdf 目前未授權公開取用 | 3.82 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。