A Simple Method for Chinese Video OCR and Its Application to Question Answering

Abstract

Captions in videos contain valuable information for video retrieval. Although texts in captions can be obtained easily in the new image compression formats like MPEG2, there still are many video programs encoded in older formats. Thus, video OCR is indispensable for content-based video retrieval. This paper proposes a simple video OCR method for Chinese captions, including image capturing, caption region deciding, background removing, character segmentation, OCR and post-processing. We employed Discovery Channel films as training and testing corpus. In an outside test, the accuracy of the video OCR was 84.1%. The hardware used in the experiment consisted of a computer with a P4-1.7G CPU, 256MB RAM and a 40G, 7200rpm hard disk. On average, it took 29 minutes and 11 seconds to process a film 495MB in size. We also applied the results of video OCR to video retrieval and question answering.