1. INTRODUCTION
Many companies such as marketing research and advertisement are interested in identifying commercial breaks from live TV streams. They may want to verify if a TV commercial has been actually broadcasted as contracted; and they may also want to know how their competitors are conducting their advertisements. It is highly desirable to have an efficient system for automatic recognition of TV commercials and storing the recognized advertisements in a video database, which can be retrieved based on video content from the database when requested. A typical automatic TV commercial recognition system consists of the following steps. Firstly, live digital TV video data is captured and stored in a computer system. All commercial breaks are identified and segmented into commercial spots (i.e., individual commercials). Secondly, a fingerprint of each commercial spot is extracted. A commercial spot comprises a sequence of image frames. The fingerprint of a commercial spot is a representation of the entire video sequence, computed from important features from component frames. Thirdly, a new commercial spot appeared for the first time should be recognized by human, and its fingerprints and identifier, such as the key number, are stored in the video database. The key number is an identity for a commercial spot provided from advertisers or marketing agency to for commercial spot identification. Finally, new commercial spots will be compared with the commercial spots stored in the video database to see if it is similar to any known commercial spots, using their fingerprints. The last step is the most difficult one, in terms of both effective similarity matching and efficient query processing. Due to many factors introduced by different broadcasting stations, digitizing and recording methods, the same commercial video sequence could be different (for example, slightly different frame rate and color variations). On the other hand, different commercial spots may have very similar video contents. An example is shown in Figure 1. Video clips A and B are different commercial spots with very similar video sequences. Video clips A and C are the same commercial spots but have are slightly different in the last frame. Only 5 frames are shown from each sequence here. In reality, the number of different frames could be very large, for example, 15 seconds video clips. The challenge here is how to match A to C but not to B. Clearly, the answer depends on the level of difference between A and B, and the level of similarity between A and C. However, a unique and robust fingerprint plays a critical role in matching video sequences. Sample Video Sets