利用影像中隱含的場景結構引導影像填補之研究

Abstract

近十幾年來，影像填補已經成為一項在電腦圖學和電腦視覺領域中重要的技術。然而，大多數前人的方法都只利用到鄰近區域的資訊，例如: 顏色,梯度,光照。這些方法並沒有考量到更高層次的資訊，所以它們的結果常常會因為消失區域的增加而變得扭曲。近期的方法分別在進入最佳化處理之前利用偵測重複性、判別不同平面、或手動先塞給一個關鍵影像片段來試圖解決此問題。在這篇論文中，我們提出一種在進入影像填補最佳化之前先對影像做中階分析並擷取出重要結構資訊的方法。我們根據直線的延伸性,物件的封閉性和區域性的一致性作分析。然後利用擷取出的資訊來引導低階的最佳化處理使場景的結構能因此被保留住。實驗顯示，我們能成功地維持這些結構特徵，也因此能在具有結構性的影像中得到比之前的經典傑作們更加合理的結果。

Image completion is an important technique in the field of computer graphics and vision in recent years. However, most of previous methods utilized properties in the vicinity such as color, gradient and illumination. They did not consider higher level information, so the results often distorted when the missing region become large. Several works tried to solve this problem by detecting the regularity, distinguishing different planes or manually inputting a key patch before the completion process. In this thesis, we propose performing a mid-level analysis first and extracting important structure information before the optimization of image completion. We analyze line extension, closure properties of objects and compartment integrity. Then, we use this information to guide the low-level patch-based optimization so that the scene structure can be preserved. The experiments show that we succeed in maintaining these structural features and generate more reasonable results than those of previous state-of-the-art methods.