以文件組成結構進行投影片比對之研究

Abstract

本篇論文提出一影像比對的方法，可用於影像畫面與原始文件的比對。在不需要建立原始文件的背景下，擷取出文件的文字及圖片區域，建立結構描述後進行比對，使原始文件不需受限於相同背景。比對方法分為四個階段：(1)擷取影像文件內容 (2)進行文件分析，擷取影像文件的文字及圖片區域 (3)分別影像文件的文字及圖片區域建立結構描述 (4)以結構描述進行比對，計算影像文件與任一原始文件之結構描述的可信度值，以可信度值最高的原始文件作為比對結果。 本研究利用 51組、共1153張投影片，每組投影片皆有一段對應的影片，比對的進行是由影片中擷取出影片畫面，利用所提出的比對演算法，找出與影片畫面對應的原始文件，實驗結果可達到 97%的正確率，能夠克服拍攝品質不良及投影片內容組成複雜的問題。

This thesis proposes an image matching algorithm for matching up video document images against original documents. Our approach extracts the text and picture regions from the document, then build structure description to match up the original document without reconstructing the background. This makes original documents be manufactured with different background. The algorithm consists of four steps. First, the document content is segmented from the video and calibrated to the video frame size. We named the document content video document. Second, the video document is processed by document analysis, then the text and picture regions are extracted. Third, the text and picture region are used to build structure description individually. Finally, we calculate the confident value between the video document and each original document, and take the one which have the highest confident value to be the matching result. Experiments were conducted using fifty-one sets of slides and video files, and the number of all slides is 1153. We use the proposed algorithm to match up the video frame against the corresponding slide. The experimental results attain 97.4% precision rate in total slides. This shows the algorithm can be applied to the low quality video and slides with composite contents.