Captions present in video frames play an important role in understanding video content. This paper presents a fast algorithm to automatically detect captions in MPEG compressed video. It is based on statistics features of caption text's chrominance components. The paper also discusses its principle and speed-up mechanism in detail. We have successfully exploited the technique to automatically construct the pictorial catalogue, a new content representation. Experiment results show the proposed caption detection algorithm has not only the ideal accuracy 96.6% and recall 100%, but also a detection speed of faster than real time.
展开▼