A Localization Method of Video Text in TV Program Video

摘要

Video text superimposed on a TV program includes useful information for analyzing the content of the program. For example, such text may represent a person's name, the substance of an utterance, or the score of a match in some sport. This report introduces our technique of localizing video text with high accuracy. We used a machine learning framework and image features reflecting the layout of character components such as lines and points to improve detection accuracy. In an experiment on broadcast TV programs, our technique achieved a recall rate of 89.9%, precision of 88.0%, and F-value of 0.889, which is a 0.135 improvement in F-value over existing techniques.