請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/78651
標題: | 基於原始對偶演算法之光場超解析影像技術 A Primal-dual Super-resolution Algorithm for Light Field Images |
作者: | Chien-Han Hsu 徐千涵 |
指導教授: | 盧奕璋(Yi-Chang Lu) |
關鍵字: | 光場,針孔陣列,深度估測,超解析度,一階原始對偶演算法, light field,pinhole array,disparity estimation,super-resolution,first-order primal-dual, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 光場攝影技術能記錄場景中的空間及角度資訊,因此具有數位變焦的功能。使用者能藉由空間及角度的資訊重建二維影像,不需要於拍攝影像時事先將相機對焦於特定的位置。然而由於部分的空間解析度被轉換為角度資訊,所重建之二維影像解析度將大幅低於捕捉影像之相機感光器解析度。 本篇論文結合深度估測以及超解析演算法,提出了一套針對光場影像產生高解析度全對焦二維影像的完整架構。由於所採用的光場設置需要先根據深度資訊,將正確的全對焦彩色影像自光場資料提取出來,我們改良了現有的深度估測演算法,利用其產生低解析度的觀察影像作為超解析演算法的輸入,並考量光場影像的性質,設計問題的目標函數,再利用高效率的原始對偶演算法進行最佳化。此演算法保留了影像細節,同時也保存了低解析度觀察影像中的自然度。 我們將架構實測於不同的拍攝影像以及不同的超解析倍數如2 × 2,3 × 3,以及4 × 4。在一張四維影像尺寸為24 × 37 × 51 × 51 的光場影像上,三種倍數的超解析流程皆可在100 次迭代內收斂,其中3 × 3的二維影像尺寸為1056 × 1680。本篇論文中應用於超解析及深度圖修補之原始對偶演算法可彈性適用於不同函數,使用者可根據所側重之輸出影像性質設計最佳化目標。利用本篇論文提出的架構,使用者能夠由單一的四維光場影像得到高解析度的全對焦二維影像。除此之外,此架構的建立有助於分析其內部不同演算法之間的搭配,有助於針孔陣列光場研究的發展。 Light field photography is a technique which records spatial and angular information of a scene, and provides capability of digital refocusing. Users are able to reconstruct 2D images based on the angular and spatial information without choosing a specific focal point when taking photos. However, since spatial resolution is traded for angular resolution, resolution of reconstructed 2D images is far less than resolution of camera sensors. This thesis proposes a complete framework for high-resolution digital all-in-focus image processing, which integrates disparity estimation and super-resolution. Since our setup requires to extract all-in-focus color images from light field data based on depth knowledge, we propose a new disparity estimation algorithm, and use it to generate low-resolution observations for super-resolution. Considering the properties of light fields and formulating the problem into an objective function, we then optimize the problem with an efficient mathematical tool. The algorithm preserves details in high-resolution products and meanwhile keep naturalness from original observations. We have experimented on various captured data and performed super-resolution with scales of 2 × 2, 3 × 3, and 4 × 4 to validate the proposed framework. Optimizations of all scales converge in 100 iterations on a 24 ×37×51×51 4D image, where the pixel sizes of the 3×3 result is 1056×1680.With the proposed framework, users are able to obtain a high-resolution all-in-focus image from a 4D light field image. The primal-dual algorithm applied in super-resolution and disparity inpainting has a flexible structure, with which users can employ suitable objective functions according to the desired properties of output images. As demonstrated in this thesis, the proposed framework is effective and can be easily adopted to different applications in light field photography. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/78651 |
DOI: | 10.6342/NTU202004323 |
全文授權: | 有償授權 |
電子全文公開日期: | 2023-11-22 |
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 | |
---|---|---|---|
U0001-0511202017021100.pdf 目前未授權公開取用 | 43.37 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。