Hierarchical system for objectionable video detection

As a growing number of Internet users are exposed to objectionable videos, the need for preventing those videos from being spread through the Internet has increased significantly. In this paper, a novel multilevel hierarchical system for detecting objectionable videos is presented. It consists of three phases using multiple features in different temporal domains. In the first phase, early detection is performed based on hash signatures prior to the download or play of a video. Next, the real-time detection phase statistically estimates the degree to which the video may be objectionable by using single frame based features. In the third phase, posterior detection is performed based on group-of-frame features reflecting the overall characteristics of the video. Notably, the second phase is a new technique for real-time detection of objectionable video, attempted here for the first time. The proposed system not only achieves high detection performance, but also enables fast detection of a variety of real-world videos.