Objectionable Streaming Video Filtering Based on Video Classification

Abstract

Due to the popularity of the computer-sharing networks, the general public has more easy access to objectionable Internet information, which is not suitable for minors. Therefore, the methods to stop and prevent the minors to access these improper information on the Internet become an important issue. In this work, the improper information is defined as the pornographic images and video streams. However, the improper video streams usually contain both benign and objectionable video segments. We propose a two-tier SVM classifier method to decide whether a given video stream is a proper or an improper video film and the shortest time elapse required to determine. In the first tier, we adopt the traditional image filter to classify video frames and get a series image classification result, which may contain false-positive and/or false-negative classifications. From the first tier image classification result we then extract the potential video features. In the second tier, we use a second SVM classifier to classify video clips and decide whether the given video stream is benign or objectionable in the shortest time frame possible. We show that the propose methods can effectively compensate the inaccuracy caused by using the image classification results alone to classify videos and to enhance the accuracy of video stream classification.