老舊發霉影片之數位修復

Abstract

這篇論文將會介紹一個數位影片修復(video restoration)的系統，這個系統可以修復大部分老舊發霉的影片。而這個論文主要有兩個貢獻。第一個貢獻是利用機器學習(machine learning)的方法自動偵測老舊影片發霉的部份。這個機器學習方法的名字叫做AdaBoost，它主要的作用在於選擇有效的特徵(features)用來學習。第二個貢獻是提出一個新的影像補洞(video inpainting)演算法用來修補偵測出來的破洞。這個方法與傳統的方法有著很大的不同，它最基本的原理是利用光學流動(optical flow)技術來填補破洞，除此之外我們利用反覆(iteration)的計算來加強結果，並減少因為光學流動估計所產生的錯誤。在本論文中還提供三段老舊發霉影片的修復片段，以及一段影像補洞演算法的結果。

This thesis describes a video restoration framework that can handle most moldy pat-tern in an aged film. There are two key contributions. The first is a learning algorithm, based on AdaBoost, which selects a small number of critical features and yields an efficient strong classifier to detect moldy pattern in an aged film. The second is a new video inpainting algorithm which repairs the detected holes. This video inpainting al-gorithm is much different with traditional ones. It is based on the temporal informa-tion from the optical flow. Furthermore, we refine our result by iteration in order to reduce the error, which is generated by inaccurate estimation of optical flow. Three examples of aged film restoration, and one example of video inpainting are presented in this thesis.