Scene Duplicate Detection from News Videos Using Visual Features and Audio Features

In this report, we propose a method to detect scene duplicates from news videos. Scene duplicates are videos taking the same event from different perspectives. Recently, technology to deal with the large amount of news videos based on their contents is needed. Among them, we are focusing on the scene duplicate detection problem. To detect scene duplicates, we have previously proposed a region-based matching method. This previous method is robust to minor difference of perspectives. However, it was difficult to detect shots in the case of a major difference of perspectives. In order to deal with this problem, the proposed method detects scene duplicates by audio features in addition to image features. Although a detectable shot is restricted, audio features can detect scene duplicates regardless to perspectives. Therefore, we expect to accuracy to improve. Scene duplicate detection experiments were conducted using actual news videos. From the results, we confirmed the effectiveness of the proposed method.