以人因分析為基礎之立體影像觀看舒適度最佳化及 雙視角轉多視角方法

Abstract

在此篇論文中，我們提出了一個可將雙視角立體數位內容轉為多視角立體數位內容的方法。此方法是以人因實驗之結果作為基礎，對輸入之立體數位內容做深度方面的最佳化，以保證觀測者觀看時不感到視覺疲勞、頭暈等負面感受。 隨著立體電影再次在面板業間引起熱潮，兩個關鍵因素仍然抑制著立體電視蓬勃發展。第一個因素為人在觀測立體數位內容時，常感到視覺疲勞或頭暈等症狀，而另一個因素為觀測立體內容時必須配戴專用的立體眼鏡的不方便性。第一個問題已由此篇論文所敘述之方法解決，而隨著面板製造業的廠商正努力於解決第二個問題，此篇論文同時提出了可由雙視角立體數位內容轉換出多視角立體數位內容的方法，防止未來裸眼立體電視大量生產時相應的多視角立體數位內容供不應求。此套方法所涉及的相關技術包含視差計算、影像分割、深度調整法、深度傳播以及影像濾波等。 此方法各個步驟中包含了多個新提出的演算法，其中包含了高精確度的視差計算方法、加速視差計算的方法、改善經深度調整後之影片連續性的方法、避免調整深度時產生嚴重變形的方法，以及最終目的－改善觀看立體內容時視覺疲勞或不適的問題。為了檢測這些新方法的效果，我們設計了幾項實驗。針對視差計算的方法，我們與其他已公佈之方法比較精確度與速度。至於可相容於其他系統的幾種加速或改善方法，我們以自己建立的系統測試使用了這些方法的前後結果。實驗結果顯示這些新提出之方法確實有效並可改善整個立體數位內容處理系統。

In this thesis, a complete method that generates visual fatigue free multi-view 3D contents from two-view 3D contents is proposed. As a new trend of 3D arises in the display industry, two critical problems still stand in the way, visual fatigue and the inconveniency of wearing 3D glasses. The first problem is resolved by adjusting the disparity range of a given stereo content. Techniques involved in this process include stereo matching, color segmentation, clustering, depth image based rendering (DIBR), depth propagation, etc. Display manufacturers are still working on products that may overcome the second problem. In the meantime, we focus on developing two-view to multi-view conversion tools so that an abundance of multi-view contents can await the appearance of the first multi-view auto-stereoscopic display on the market. Techniques used include depth image based rendering (DIBR), disparity map pre-processing and image filtering. Experiments were designed to evaluate the effectiveness of several newly proposed algorithms. The purposes of these algorithms include generating high-quality disparity maps, accelerating the process of generating disparity maps, increasing the fluency of 3D video contents, preventing distortion after view rendering and the overall effect of reducing visual fatigue. The resulting disparity maps of our stereo matching tool were compared with other published algorithms, while other proposed techniques that can be adopted by other systems were tested by comparing the before and after results of our own system. The experiment results strongly support the benefits of these algorithms.