A robust video text extraction method based on text traversing line and stroke connectivity

Automatic video-text extraction is an important field in video content comprehension. In this paper, a robust video-text extraction method is presented which can automatically extract horizontally aligned text with different languages. First, an unsupervised paradigm based on Haar wavelets is applied to obtain candidate text region. Second, the traversing line with its aptitude spectrum is introduced and applied to create boundary for each text line. Last, traversing line of the maximum feature value in each refined text region is employed to seed key-points in strokes, from which region growth is performed to create binary text image. Experiments conducted with a variety of video sources show that the method is robust to text of various colors, fonts, sizes in complex image, and performs better than conventional methods.