應用於三維立體視訊之低複雜度深度圖壓縮演算法

Abstract

MPEG 自由視角電視標準 (FTV) 之架構，包含深度資訊；若深度資訊經由現有的壓縮標準 (MVC) 壓縮，則會產生兩個問題：一、在物體邊界處容易造成深度資訊之失真，並影響合成虛擬影像之主、客觀品質；二、MVC 有兩種預測方式，分別為Intra與 Inter 之預測方式，由於壓縮後深度圖與自然影像之資料量相差懸殊，即使利用運算複雜度較 Intra 高的 Inter 預測方式也顯得沒有壓縮效率。本論文闡釋一個利用深度圖之特性，設計簡單且有效率之深度圖壓縮演算法。

本篇論文將分析單張深度圖之特性並使用眾數演算法挑選需要壓縮之深度資訊。其次，將需要壓縮之深度資訊以預測模式之方式，估測該深度資訊，企圖減少需壓縮之資訊量。最後，經由壓縮重建後的深度圖使用線性權重之過濾器以修補深度圖。實驗結果顯示，本文提出的方法在不影響合成結果之PSNR情況下，其需要的網路頻寬較 MVC 少4~13%，且運算複雜度也較 MVC 之 Intra 預測方式少 5% ~ 15%。

In the framework of MPEG Free-viewpoint Television (FTV), the depth information is included. Once the depth information is coded by Multi-view Video Coding (MVC), two kinds of problems may occur. The first one is the depth errors around the boundaries of objects, which might affect the quality of rendered virtual view. The other is that inter prediction cannot provide high coding efficiency for depth map. The size of compressed image is much higher than depth map. Therefore, a simple and efficient coding method is proposed based on the characteristic of the depth map. The proposed method analyzes the characteristic of a single depth map and uses mode determination to calculate the depth information. Then, the depth information is predicted by mode decision to increase the compression efficiency and linear weighted filters are used to refine the reconstructed depth map. The experiments show our methods save 4-13% bandwidth with the same PSNR and the computation complexity is less than the one needed for Intra prediction.