Acoustic Pornography Recognition Using Recurrent Neural Network

This research proposes a pornography recognition model using audio features utilizing recurrent neural network. The proposed work is totally different from most of the previous attempts on pornography recognition using visual contents of the nudity images or videos. The importance of using sounds for pornography recognition arises in the cases in which the visual features are not adequately informative of the nudity contents such as dark scenes, and scenes with the covered bodies. Acoustic recognition model selects the finest signal representation by feature extraction of Mel-Frequency Cepstrum Coefficients (MFCCs) after been pre-processed for noise reduction. Extracted features are fed into a network with long short-term memory (LSTM) cells. The LSTM cells have the capability to solve problems associated with temporal dependencies and require learning long-term and solve the vanishing gradient problems associated with Recurrent Neural Network (RNN). A dataset of porno-sound samples has been created based on an existing pornography video dataset. 800 data samples were used including positive and negative samples. The experimental results confirm the feasibility of the proposed acoustic-driven approach by demonstrating an accuracy of 86.50%, and F-score of 86.83%, in the task of pornography recognition.