A Heuristic Approach to Caption Enhancement for Effective Video OCR

摘要

We present a heuristic approach to enhancing speech synchronized captions for video OCR, as a pre-process for subsequent tasks of multimedia indexing, segmentation and retrieval. We use a bi-search based caption transition detection method to improve efficiency, which adopts a simple heuristics that the same caption content usually lasts for a period for stable viewing. We propose a combination of color mask, changing mask and region mask to perform caption enhancement based on the discriminative characteristics of captions and backgrounds. Elaborate enhancement on individual characters is further used to remove small background residues. OCR experiments show that our caption enhancement approach brings a high character accuracy of 89.24%.