Text localization in video/scene images using Kirsch Directional Masks

Text plays its vital role in visual content analysis and understanding. Videos contain text with diversity in its text patterns and complex backgrounds. In this paper, we propose an approach based on compass operator for detecting the edges. We obtain the edge maps by convolving the Kirsch Directional Masks along eight different directions for the preprocessed video frame. The resultant images are binarized and the edge maps are analyzed and concatenated to obtain a single binarized image. Further, we employ geometrical rules and shape properties to eliminate the false positives. Then, the text regions are identified using connected component analysis and bounding boxes are placed for the detected text regions so that the text gets localized. Experimental results obtained on the standard datasets reveal that the proposed method effectively detect and localize texts of various sizes, fonts and colors in videos and scene images.