Video Text Detection Algorithm Research Based On Maximally Stable Extremal Regions

With the development of the Internet and multimedia, large amount of videos are available on Internet, so video indexing gradually becomes an indispensable task in human’s daily life. Today, most video indexing is still done by manual tag. As we all know, it is usually not accurate. However, video texts including captions and scene texts provide most direct information, and can better express the theme of videos. Accordingly, semantic based video content analysis has become a hot field of multimedia research efforts.Text detection is a challenging problem, since video image background is generally complex and its subtitles often appear color bleeding, fuzzy boundaries and low contrast due to video lossy compression or low resolution. In this paper, we propose a robust framework to solve these problems. Firstly, we exploit the gradient amplitude map (GAM) to enhance the input image edge, which can overcome the color bleeding and fuzzy boundaries; Secondly, a two directions morphological filtering is developed to filter part of background and highlight the text contrast; Thirdly, Maximally Stable Extremal Region (MSER) is applied as a region detector to detect color of bi-polarity text region, and we use the mean intensity of the regions which detected by MSER as the Graph Cuts’label set, HSI color space H, S, I three-channel Euclidean distance as the Graph Cuts smooth term, to get optimum segmentation; Finally, we group them to text-line with the geometric characteristics of the text, then multi-frame verification and some heuristic rules are used to eliminate non-text region.We demonstrate our results under some challenging videos, and the results prove our detection framework is more robust.