Obscenity Detection in Videos Through a Sequential ConvNet Pipeline Classifier

Abstract

The amount of pornographic material available on the Web is staggering. This content is available freely on the Internet and without any restrictions which pose a threat to minors as they might be introduced to such content, which is harmful to their mental health for the long term and also there is huge threats of pornographic content generation to the celebrities, known figures due to invention of deep fakes. There are many software that blocks access to the visually disturbing sites that contains obscene, child pornography, or material “harmful” to minors but only a few software analysis motion content of the videos and mostly only image features are analyzed. To tackle this problem, we propose a frame sequence ConvNet pipeline using ResNet-18 for features extraction and analyzing <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${N}$ </tex-math></inline-formula> frame feature map using the proposed ConvNet for the frame sequence classification, therefore implicitly encapsulating the motion information by encoding changes in the ResNet output feature vector, which achieves a state-of-the-art accuracy of 98.25% in classifying pornographic videos of the Pornography-800 data set and state-of-the-art accuracy of 97.15% in classifying videos of the Pornography-2k data set.