CNN-based Commercial Detection in TV Broadcasting

TV is an important advertising media. Information of a piece of TV commercial, such as broadcasting time, the duration, the casting and etc., may reflect the business value of the host company of this commercial. An automatic commercial detection system is needed for third-party by business analysis. Previous works about TV commercial detection just apply to one kind of video, like news, sports. And those methods detect commercials at frame level, which need expensive computing cost. In this paper, we design and implement an automatic commercial detection system for TV broadcasting. This system works at shot level and detects commercials in streaming videos, including TV broadcasting and online videos. It consists of two modules, the shot boundary detection module and the shot classification module. We crawl actual broadcasting videos and split them into shots, classify these shots into two classes, commercial and non-commercial. Then, we extract shot features with deep convolutional neural network, and train a support vector machine classifier to complete shot classification. Combining the state-of-the-art convolutional neural network with traditional machine learning techniques, our system can handle various program types and indiscernible commercials, and get precision 93% and recall 95% in realistic TV programs. Except the TV broadcasting, our system works in other video media like online videos.