A robust system for text extraction in video

This paper presents a novel system to extract caption text in video. Firstly, text regions are detected primarily with emphasis on the recall rate. Then a multiple stage verification scheme is adopted to discard false alarms and boost the precision rate. Secondly, a text polarity estimation algorithm is provided. Based on it, multiple frame enhancement is conducted to strengthen the contrast between text and its background. Finally, a connected component filtering method is proposed to generate clear segmentation results and improve recognition performance. Experimental results confirm that the proposed system is robust and efficient.