PDF文件影片關鍵畫面擷取之研究

Abstract

本篇論文提出一個偵測影片關鍵畫面的方法，可自動擷取使用PDF文件為教材之影片發生換頁的畫面。偵測方法分為兩階段，第一階段找出可能發生換頁之處，利用P、B畫面運動向量方向、位移量及所佔比例，找出換頁的位置並過濾不為換頁的畫面，如：移動程度較小、區域移動、移動偏於某一方向的畫面；第二階段為去除多餘畫面，利用畫面DC值直方圖的熵找出畫面顯著區域，再藉由其間的差異判斷其是否為換頁或其他多餘畫面，如：縮放、移動距離較大的畫面，經過上述兩階段存留下來的畫面即為關鍵畫面。本研究分別針對使用不同參數設定之影片、使用不同類型的PDF文件、不同移動速率及動作的PDF文件進行偵測分析，實驗結果證明所提出的方法不論在處理速度及結果均有不錯之效果。

This thesis proposes an automatic key-frame extraction method for detecting PDF document change in videos. The algorithm consists two major steps. First, candidate key-frames are extracted by analyzing angle, magnitude and proportion of motion vectors. Second, the entropy of the DC-value histogram is computed in order to find salient objects as main information of a frame. Redundant frames are removed from candidate key-frames by noting changes in higher entropy histogram profiles of a frame. From above methods the true key-frame of the particular shot are extracted. Experiments were conducted using the video clips which involve different argument sets, different contents, motion rate and action of PDF document. The experimental results demonstrate that our method is able to detect key-frames accurately in reasonable amount of time.