請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/54680
標題: | 針對深度為基礎的即時影像應用之光場處理系統 Light-field Processing System for Depth-based Real-time Video Applications |
作者: | Hong-Hui Chen 陳泓輝 |
指導教授: | 陳良基(Liang-Gee Chen) |
關鍵字: | 多視角,光場,透鏡陣列,相機陣列,重新對焦,深度搜尋,晶片, multi-view,light field,lens array,camera array,refocus,depth search,chip, |
出版年 : | 2015 |
學位: | 博士 |
摘要: | 本篇論文探討多視角光場資訊的信號處理,首先介紹了以深度為基礎的多種影像處理與應用系統,其中光場信號處理相關的系統是本論文探討的主軸,從輸入的多視角光場信號中,採用適當的深度搜尋演算法後,即可得到深度資訊供後續應用,例如:影像重新對焦、人體姿態分析與移動物體追蹤,都可以藉由深度資訊的輔助,達到更佳的效果與準確度;我們改造一台單眼相機的機身並為其裝置上5x5的透鏡陣列,擷取多視角的光場資訊後,經過適當的校正過程後,以軟體演算法的實作證明了深度輸出的可行性並檢驗深度資訊的品質,我們發現以多視角的光場資訊來做深度搜尋,深度輸出的品質可以明顯優於現有一些只用兩個視角來做深度搜尋的演算法;接著我們的目標是把這樣的演算法應用在即時性的1080p高解析度視訊處理上,以往的實作成果多集中在單張影像的加速處理上,少數針對視訊速率處理的作品,規格等級停留在雙鏡頭720p/30fps,不及未來運用所需,我們的晶片實作,大幅拉高系統規格至五個鏡頭1080p/30fps,證明硬體架構的可行性,以現今的晶片製作水準,的確可以在一定的功耗與晶片面積要求下,實作出我們設定的硬體規格,我們採取了許多硬體實作上平行化的技巧,同時也在演算法的輸出品質與硬體的複雜度之間做了有效的取捨,克服了系統記憶體頻寬、運算量與功耗調適性種種的挑戰後,最終可以收斂到一個有效的設計,讓我們採用的深度搜尋演算法,未來可以應用在視訊速率的光場信號處理系統中,大幅開拓了光場信號處理在未來的應用可能性。 This thesis discusses the signal processing of multi-view light field. First we introduce several depth-based image processing and application systems. Among them, the light-field based systems are the main topic of this thesis. Applying depth search algorithm to the input light field, the depth information is extracted for the following different applications, for example, image refocus, skeleton analysis of human body, and tracking of moving objects. With the depth information, many algorithms and applications can achieve better performance and accuracy. We modify a digital single-lens reflex (DSLR) camera and put a five by five lens array on it. After capturing the raw light field, we apply suitable calibration process on the light field. We implement software version of the adopted depth searching algorithm to prove that the acquired light field is applicable to generate an adequate depth map with satisfying quality. We have confirmed that the resulting depth map quality from our captured light field outperforms the depth map from stereo depth searching algorithm (by using only two cameras). Then we set up the goal to apply the algorithm on a system capable of dealing with 1080p high definition images at video rate. Plenty of the previous works focus their effort on accelerating the processing on single still image. Several works targeting the light-field processing at video rate deliver their specification only up to 720p/30fps when generating the depth map. This specification cannot fulfill the need for future application. Our chip implementation supports processing video data from five cameras up to 1080p resolution at 30fps. We have proved that the proposed architecture is practical under current CMOS fabrication technology. The chip is realized under constrained power and area budget. We adopt the techniques of hardware processing parallelism and made tradeoff between the depth map quality from the searching algorithm and the hardware complexity. By overcoming the difficulties on system memory bandwidth, computational load, and power scalability, we have converged to an effective chip design. With this result, we expect that the adopted depth searching algorithm can be used in future light-field processing occasions that require video rate capability. A broader adoption of the light-field processing in different future applications is foreseeable. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/54680 |
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 | 52.68 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。