A Video Text Extraction Approach Based On Color Clustering And Connected Component Analysis

Superimposed texts in videos bring important semantic clues for content understanding. Thus, it is a key technology for videos'content understanding, indexing and retrieval to extract text from videos effectively. Commercial OCR system made a great success and the technology of text recognition from binary image is popular. However, most superimposed texts surrounded by complex background or accompanied with high noises in videos pose great challenges to text recognition system. Applying a proper text extraction method to make OCR system recognize video text images correctly is the main topic of this thesis.Considering the problems of complex background and high noise, we propose a video text extraction approach based on color clustering and connected component analysis. After text detection and location, the method employs both color and connected component as the main features to extract the texts from video images while removing background by connected component filling, to produce clear binary images for OCR system.The text extraction approach has been used in the video text extraction system, which implemented the whole process from inputting videos to producing the final result of OCR system. Experimental results show the robustness and effectiveness of the proposed method.