A comprehensive video text localization system based on stroke filter

Texts in images and videos are a significant source of high-level semantics for indexing and retrieval. But the localization and extraction of text are always interfered by complex backgrounds with strong edge and texture clutter. In this paper, we propose a comprehensive text localization system which combines several popular methods. First, we extract the stroke features of text using improved stroke filter. Then we employ global thresholding and projection profile to localize the candidate text regions on the stroke map. In the end, based on the stroke statistical features, we use SVM classifier to refine the text regions. The experimental results show that the system is robust to the text size and complex background.