Text Localization and Detection for News Video
Yu Song, Wenhong Wang · 2009
Texts presented in video can provide important semantic information. In this paper, we propose an algorithm to detect texts from news videos. First, SPAC (spatial autocorrelation method) is used to determine the degree of texture rough-detail of video frames and determine candidate text regions based on texture rough-detail. Then Sobel operator is used to extract edges of candidate text regions, and filter text regions of fall short of conditions based on prior knowledge. At last, we can separate texts from complex background though image binarization processing. This approach can rapidly detect various colored text regions, and reduce computation. Experiments show that the algorithm can detect text regions simply and effectively.