Pornographic Video Classification Using Convolutional Neural Network and Gated recurrent unit (GRU)

Abstract

Changes and advancement in many fields become eminent in the world where we live. In the last ten years, there has been a high coverage and availability of internet connection. These the advancement of technology and emerging of advanced computing platform bring a lot of advantages and negative influences on our community. From those negative threats, one that harshly attacks the youth’s productive class of the population is the videos with pornographic contents. According to the annual statistics released by PornHub videos, 64 million people visited PornHub every day. This is the noticeable number. This will lead youths to exercise unsafe sexual behavior and of course they are exposed to sexually transmitted diseases like HIV. In this study we have proposed an automated classification of the pornographic videos. A combined effect of CNN pretrained models along with GRU has been employed to tackle this problem. The CNN pretrained model such as EfficientNet has been used for relevant feature extraction. Where the sequential learner bidirectional GRU is responsible for detecting the instance of video frames as porn video content or not. To evaluate the proposed model a publicly available 2K NPDI dataset from the university of Campinas, Brazil and applied has been used. A preliminary preprocessing steps such as normalization and cleaning has been applied on the dataset. We have used EfficientNet as a fine-tuned feature extractor in order to extract important features from the frames of the video and then the sequential information from the frame is learnt by DB-GRU network. In this DB-GRU network multiple layers are stacked together in both forward and backward pass in order to increase its depth and get good accuracy. Beside this various parameter optimization has been applied to increase the accuracy and performance of the proposed model. Following this, the experimental evaluation has showed a significant result of 99.68%. This result has improved by 0.68% and there is an improvement in efficiency in training and testing when compared to previous attempts on similar datasets. Various visualization methods have been used to present the result in more interpretable and human interpretable way. This will help readers to reproduce our work. Finally, we have tried to show the real time application of the proposed system by integrating and deploying the model, which has been already developed in this study along with web-based API. Thus, any user can scan to detect early whether a pornographic content present in a certain video stream or not