A Real-Time And Robust System For Commercials Detection

With the constant innovation and progress of multimedia technology,digital media is becoming very important in people’s life. Advertising as a very important means of business promotion is an important part of multimedia applications.Television advertisement as a part of TV programs, its scale expands continuously. In order to expand the influence and popularity of products, merchants constantly advertise on television programs, particularly in the middle of the prime-time and hit show. But for TV audits, commercials will affect they completely watch TV programs. In addition, the high quality commercials will play a positive role in promoting consumption. But false advertisements bring immeasurable harm to the society and the public and seriously affect the image of the media and the credibility of the society. Therefore, the commercials detection from television is urgently needed to solve.This paper proposes a framework which is real time and robust to commercials copy detection. The framework consists of two parts:1. Commercials copy detection based on audio match; 2. Precisely localization of the boundary based on the image features. The first part includes the offline and online processing. Offline processing mainly deals with commercials, extracting audio from advertisements and applies short time Fourier transform to the audio to obtain the spectrograms. Then we extract feature descriptors form the spectrograms and build a feature database. Online processing mainly deals with the test video, and find a match from the database.The second part first merges the commercial clips detected in the first part. And we use a strategy to smooth the mistaken segments inside the advertisement and normal video. Finally, we use the image features to find the boundary between the advertisements and normal video precisely. The experimental results show that the accuracy of the final result is as high as 99% which is very close to the actual situation. The run time will not increase withe the size of the database increases, and is close to real time.