Multi-scale Text Extraction Algorithm In Video Sequence

Nowadays, video is the most popular media form in airwaves, Internet and wireless network. For the sake of that the users can find their interested information as soon as possible, many researchers devote themselves to video retrieval. In video sequence, the emedded text provides important information for high-level semantic analysis in multi-media data. It is a vital clue for video indexing and summarization. Almost all the research of video retrieval was started with text detection and extraction. Therefore, how to localize and extract text from video data under complex-background is one of the hot topics.This paper aims to make a deep study into several key techniques on an automatic edge-based multi-scale text extraction algorithm in video frame sequences. It analyzes the features which are usually used in text location and extraction, then, selects the features which are effective in both English and Chinese and takes them as entrance of the research. Firstly, the characters of the video frame sequences is made full use of to reduce the influence of the complex-background by using multi-scale integration. Secondly, in order to enlarge the detection region of the text font, a multi-scale transform processing algorithm is proposed. Thirdly, text extraction is carried out by local adaptive thresholding binarization and inward filling, which is more efficient on the segmentation of complex-background text. Finally, the binary image obtained from text extraction was delivered to OCR software for recognition.Experimental results show that the proposed method has satisfactory performance on both English and Chinese. It especially worked well on complex-background text localization and text extraction.