DVRGNet: an efficient network for extracting obscenity from multimedia content

The availability of adult content on the internet through images or videos is easily accessible to minors and adults as well. In addition, this type of content may lead to poor mental health, sexism and objectification, and sexual violence. Therefore, It is extremely important to detect and classify pornographic content. In this paper, DVRGNet, a hierarchical CNN framework for the detection and classification of obscene content from videos is proposed. The proposed framework incorporates motion data and the capture of motion movement to deal with the problem of mapping skin exposure to pornographic content. DVRGNet is a network that leverages DenseNet, VGGNet, ResNet, and GoogLeNet for feature extraction. This network includes different fusions of various sub-networks, which can be seen as diverse tiers of neurons in human brains. The framework also incorporates a 5-layer Bi-LSTM-based classification of obscenity from videos. The proposed framework makes better use of automated pornography detection through computational intelligence architectures. Furthermore, the fusion of these four networks strengthens feature propagation by reducing the vanishing gradient problem. Extensive experiments are conducted on Pornography-2K and Pornography-800 datasets to validate the effectiveness of the proposed framework. The proposed framework achieves an accuracy of 99.42% on the Pronography-2K and 99.04% on the Pornography-800 datasets. An ablation study is also conducted to demonstrate the performance of proposed framework.