Video text binarization using connected component level non-text filtering

Abstract

In this paper, we present a video text binarization algorithm using connected component (CC) level filtering. We develop a three-step algorithm: estimate the polarity of text, extract text/non-text CCs with a retinex filtering method, and filter out non-text components. The proposed non-text CCs filtering is considering CCs stroke width and contrast. We conducted experiments on subtitles in TV programs and movies. Experimental results showed that the method outperforms conventional methods in terms of a pixel-level accuracy and a character/word recognition rate.