Classification of Video Recaptured from Display Device

The prevention from unauthorized recapturing of screen is an important issue in multimedia security. In this study, we attempt to detect illegally created videos captured from display devices by analyzing unnatural signals contained in the videos. The proposed approach applies a convolutional deep neural network (CNN) for the classification. In order to reduce the computational costs, some frames are sampled from a target video, and are checked whether they are captured. In the training process, each frame sampled from captured/natural videos is partitioned into small patches, and a CNN model is trained by using the patches. The final decision is determined from the classification results at each frame. We conducted experiments to evaluate the classification accuracy and its dependency on camera devices. It is confirmed that we can classify captured and natural videos with high probability under our experimental conditions. When a same camera device is used for recording both original and recaptured videos, the classification accuracy is decreased from the case of different devices.