基於多線索與3D觀看舒適度增強之2D至3D轉換系統

在這個3D技術發展已達一定水準的年代，各式的3D顯示產品蓬勃發展，卻未見3D普及深入一般生活，其原因在於，3D影片的來源還不充足。如能將現有的2D影像/視訊轉換為3D影像/視訊將可彌補此問題。目前常見的轉換技術是利用2D影像中具有距離資訊的線索去估測出2D影像的深度資訊並利用DIBR技術產生虛擬的多視角3D影像，然而2D影像存在著多樣的資訊，所以深度估測演算法對於來源的適應性便非常重要。本論文提出多線索並結合可靠度機制深度圖選擇來提升整個系統對於多樣來源影像深度估測的適應性。本論文將來源影像分類並針對各類別的特性進行分析，提出有效的多線索深度估測演算法，其中使用了消失點、對焦、主要直線分布…等多種線索進行深度估測，針對對焦線索，我們提了一個新的對焦線索之深度估測演算法，利用計算物件邊界強度來進行深度估測，而在消失點的深度估測中，我們結合了即時的消失點估測與物件分類產生較佳品質的深度，最後我們提出了利用主要直線分布來進行低線索依靠度的深度估測。在藉由多線索進行估測出多張深度後，提出之可靠度分析系統會藉由分析各線索的可靠度來決策最佳深度。另外，因序列深度間的穩定度直接的影響3D影像觀賞的舒適度。為了確保深度圖的觀看品質與舒適度，本論文不同於一般僅討論單張靜態影像深度估測的演算法，本論文提出序列深度圖的穩定機制，將序列相鄰深度間的關係性納入考量以達到穩定序列深度的目的。在實驗結果的部分，由客觀分析及主觀分析結果，我們提出的多線索深度估測演算法針對不具有相同線索的來源可以維持一定品質的估測深度圖，並藉由深度穩定系統可以更進一步提升深度圖的效果與觀看舒適度。

關鍵字

2Dto3D ；深度估測；多線索；可靠度；穩定增進；適應性演算法

並列摘要

In recent years, 3D technology becomes a mature consumer technology. There are several 3D displayers in the market. However, 3D is still not general in our life. The reason is the lack of 3D video content. To convert the existing 2D image/video into 3D is one way to solve this problem. The 2D to 3D conversion technique is developing to estimate the depth information from the useful distance information, like vanishing point, foggy and focus, in the 2D image/video. From the estimated depth information and original 2D image/video, DIBR technique can generate the virtual multi-view for 3D viewing. Since the distance information of 2D images/videos is too many, hence using a specific cue to estimate depth is not suitable. Therefore, an adaptive depth estimation algorithm is necessary in the depth estimation algorithms by using multi-cue. This thesis combines multi-cue depth estimation and depth reliability analyzing system to reach an adaptive algorithm. First, 2D images/videos are classified into 3 types, long shot, close-up and others. For each type of 2D images/videos, thesis proposed effective depth estimation algorithms by using different cues. Vanishing point cue is for long shot. Focus cue is for close-up. Main lines distribution (MLsD) cue is for others. This thesis uses the strength of object boundary to model the blur caused by focus and defocus and estimate depth map from blur information. For long shot type 2D images/videos, this thesis combines real-time vanishing point detection and depth assignment by using object regions to generate a fine depth map. For others type 2D images/videos, this thesis uses common MLsD to generate depth map from predicted depth map set. According to the reliability score of each cue, proposed depth reliability analyzing system will choice the best depths which are estimated from 3 different cues. Since the stability of generated depth sequence affects the 3D watching experience and comfortable. For viewing comfortable, the scene change information of neighbor depths and input 2D images in sequences are considered in proposed system to preserve the stable of the generated depth map. From the experiment result, the objective assessing result shows that proposed depth estimation algorithm can provide a high quality depth map. The depth stability enhancing system can promote the 3D virtual image for viewing comfortable.