數位視訊中自動偵測字幕之研究

Abstract

影像中的文字可以幫助我們瞭解影像所要傳達的資訊，若能夠偵測出影像中的文字，可以用來幫助我們做大量影像資料庫的瀏覽、索引等工作。本論文提出一套用於視訊畫面，自動地做字幕文字偵測的方法，以得到視訊畫面中內容的資訊。 利用影片中字幕文字的邊緣特色，本論文提出一套可同時偵測水平與垂直文字區域的方法。為了克服文字區域邊界的文字可能因特徵不明顯而不容易被偵測出來的情形，我們也提出了一個根據區域邊界邊緣能量的大小向外擴張一個區塊的方法，以偵測出完整的文字區域。在偵測出可能的文字區域後，根據文字區域具有較高的水平、垂直邊緣能量與較高的邊緣密度，利用這兩樣特色，我們採用一個模糊推論系統來驗證偵測出的可能文字區域，以去除偵測出的雜訊區域。實驗的結果顯示出我們的方法可以有效且完整的偵測出文字區域。

The captions in a video frame can help us to understand the information that image carried. If the captions in a video can be detected, it can help us to make video annotation and indexing. In this thesis, an automatic caption localization method is proposed to get the information about the video content. Using the edge property of captions in a video frame, we propose a method to locate the horizontal and vertical text areas simultaneously. In order to solve the problem that some characters without clear edge feature in text area boundary may not been located. We also propose a method to extend one text block according to the edge energy at the boundary of the text area to locate the whole text area. After locating the candidate text areas, according two characteristics of text area of higher horizontal and vertical edge energy and higher edge density, we adopt a fuzzy inference system to verify candidate text areas to filter out false detected noise areas. Experimental results have shown the proposed approach is efficient and accurate in text area detection.