Automatic Video Text Localization and Recognition

Text in videos contains much semantic information that can be used for video indexing and summarization. In this paper, we design an integrated algorithm of locating horizontal text based on corner point detection and color clustering. First, we get candidate text regions by using the method based on corner point detection, and then identify candidate text regions and refine the bounding boxes by color clustering. Both the precision and recall rate of the new localization method are improved, and the processing time of the new method is less. On the aspect of locating accuracy, the new method gives tighter bounding boxes. We finally enhance the quality of the detected text region by multi-frame averaging and local thresholding. Our method can handle multi-language video text with complex background including a great range of font sizes and styles. The results after above steps can be directly processed by OCR system.