International Journal of Computer Science Issues

More than a traditional journal...

The digital age has ushered in a plethora of ways for video recapture and video tampering. Subsequently, digital video forensics has become increasingly important, in which recaptured video detection is one branch. The applications are not limited for illegal video copies detection in professional cinematography and home entertainment, and surveillance video authentication in crime scene investigation, but also being able to detect recaptured videos will enhance the robot vision and add more intelligence to security systems such as face authentication systems, by enabling them to detect live scene from reprojected one. Furthermore embedded in web, monitoring systems may provide additional tools for protection and administration of video contents which would otherwise have cost thousands of man-hours for manual screening. In this paper, an automated movie piracy detection mechanism based on multiple feature descriptors is proposed. The proposed method uses combinations of lowlevel features including amount of blur, noise, color moments and texture patterns of video frames. To demonstrate the accuracy and efficiency of the proposed method, we maintained a video dataset comprised of videos obtained at different resolutions and different shutter speeds. In order to compare our proposed method with existing state of the art, we used the same video database used in [22]. For practical purposes, videos in dataset is composed of different durations (from 30 seconds to 15 minutes approximately) and different categories including sports, educational, movies, TV commercials and animated. Deviated from [22] we have additionally included surveillance videos to the database as well. In order to obtain a recapture video database, videos were recaptured in an artificially lit room with fine tuned controllable settings. A special setup was used to ensure that recaptured videos are of high quality and they cannot be distinguished by naked eye. Extracted features are used to train different Support Vector Machines (SVMs) and a feed forward back propagation neural network. The experimental results show that our method uses a reduced number of feature dimensions and exhibits greater robustness as well as greater accuracy compared current state of the art [20] in identification of the recaptured videos. The method is capable to generalize the approach to both high quality videos as well as for the surveillance video sequences with low resolution. Therefore the proposed architecture provides an efficient and flexible solution for video piracy detection.