Fast pornographic video detection using Deep Learning

Abstract

The recent rapid development of internet technology and applications leads to the booming of videos uploaded to and shared on the internet. However, some of them may contain impermissible content, especially pornographic videos, for the viewers. This problem raises a vast challenge in video filtering for many input videos. Concerning this matter, we introduce our system in which a key frame extraction method will be applied to the input video at the very first step. Subsequently, the Tensorflow object detection API is in charge of detecting and cropping any existing person in these frames. A Convolutional Neural Network (CNN) model then takes the cropped images and classifies them as pornography or not. The video is finally is marked valid for publishing according if the number of adult frames is below a threshold. Our experiments show that the proposed system can process videos much faster than human do while the accuracy is around 90% which can be meaningful to assist people in the task of video filtering.