A Novel Pornographic Visual Content Classifier based on Sensitive Object Detection

Abstract

With the increasing amount of pornography being uploaded on the Internet today, arises the need to detect and block such pornographic websites, especially in Eastern cultural countries. Studying pornographic images and videos, show that explicit sensitive objects are typically one of the main characteristics portraying the unique aspect of pornography content. This paper proposed a classification method on pornographic visual content, which involved detecting sensitive objects using object detection algorithms. Initially, an object detection model is used to identify sensitive objects on visual content. The detection results are then used as high-level features combined with two other high-level features including skin body and human presence information. These high-level features finally are fed into a fusion Support Vector Machine (SVM) model, thus draw the eventual decision. Based on 800 videos from the NDPI-800 dataset and the 50.000 manually collected images, the evaluation results show that our proposed approach achieved 94.06% and 94.88% in Accuracy respectively, which can be compared with the cutting-edge pornographic classification methods. In addition, a pornographic alerting and blocking extension is developed for Google Chrome to prove the proposed architecture's effectiveness and capability. Working with 200 websites, the extension achieved an outstanding result, which is 99.50% Accuracy in classification.

Joint inter-intra representation learning for pornographic video classification

“…While the object-based approaches can ensure the right prediction in most cases, the strong resemblance between sexual objects with common items in some special cases or viewpoints (such as dildo and sausage) makes it difficult to make the right prediction. In our previous studies that focus on identifying sexual objects and organs on object-based approach [18]- [20], we labeled four sexual organs male/female genitals, female breast, and anus with polygon mask for both object detection and instance segmentation tasks. With the labeled dataset, we not only developed a sexual object detector based on mask R-CNN but also utilized the training strategy with two steps learning that helps the detector overcome the false positive prediction on sexual objects, thus enhancing the performance of recognizing and classifying pornography content.…”

“…However, these methods predominantly used the extracted key-frames that NPDI's author provided feeding to their model, rather than learning the representation throughout of the video that limited the model's performance. In our previous experiments on pornography videos [18]- [20], we extracted key point frames throughout the whole videos of NPDI instead of using provided key-frames, as we believe it comes with a better result in precision.…”