以獨立成份分析法進行影片文字萃取之研究

Abstract

本論文針對影片中的字幕提出一個創新的文字萃取方法，在本研究中利用獨立成分份分析法(Independent Component Analysis, ICA)來分離影像中的文字區及非文字區。為了符合ICA的處理程序，本研究提出了3種資料的表示方法，其中包含針對連續影像的「序列特徵表示法」，以及針對單張影像的「rgb特徵表示法」與「高次特徵表示法」。為了分析本方法的強韌性(robustness)，於研究中採用數種不同的字幕型態影片做為測試，並且進階的改變字幕的字型、顏色、尺寸以及排列方式。實驗結果顯示，本研究所提出的方法是一個強韌性佳的文字萃取方法，比一般傳統的方法更具系統性，並且不需要複雜的參數調整便能有效率且正確的將字幕從背景中分離出來。

In this thesis, a new text extraction method is proposed for digital video. The text in the video frames is extracted by means of the independent component analysis (ICA). There are three representations of the image frames are introduced to fit the need of the ICA procedure. The first is the sequence representation for a sequence of image frames. The second and the third are the rgb representation and the high order representation for a single image frame respectively. The font, color, scale, and orientation of the subtitles in the sample videos are changed to evaluate the performance of our method. The experimental results show that the method we developed in this thesis is robust and systematic. It is very effective and accurate for the text extraction in videos, without any complex adjustment of the parameters.