PPCensor: Architecture for real-time pornography detection in video streaming

Pornographic content classification using deep-learning

“…Increasing the -value favours a lower recall value, hence increasing the overall mAP score 3.81% from 59.82% to 63.63%. This is comparable to the mAP score of 63.5% achieved by Mallmann's Private Part Detector (PPCensor) [19]; however, the latter only discriminated between 4 classes compared to the 9 classes distinguished here. Besides detecting more than twice as many classes, our framework also displays an average throughput of 29.2 frames per second (FPS) on a Nvidia RTX2080-Ti GPU, roughly 2.7 times faster than the PPCensor which reportedly performed at 10.86FPS.…”

“…An additional benign class was used in the multi-class MobileNet to represent the possible background. Mallmann et al [19] propose an architecture that addresses pornography content detection entirely through object detection. They manually label 52,215 object instances to create their Private Parts Object (PPO) dataset which is used to fine tune a faster R-CNN Inception architecture for real-time object detection.…”

“…The three additional categories include cartoons and drawings depicting children engaged in sexual activity, benign exploitation of children, and adult pornography. The current array of deep learning tools aimed at the detection of pornographic content [11,19,29,31] are limited to flagging sensitive content and do not distinguish between different degrees of pornographic material. Thus, LEAs do not have the means to automatically categorise and organise the sensitive content seized from the perpetrators.…”

“…Thus, developing algorithms that detect sexual body parts provide LEAs with tools which not only detect pornographic content but also provide descriptors justifying the decision. We improve on Mallmann's private part detector [19] by presenting a solution that detects twice as many object classes with a 270% increase in image throughput. In addition, we extend this model to also provide a description to image documents that would help in ranking the degree of pornographic explicitness of a file.…”

A Video Summarization Approach to Speed-up the Analysis of Child Sexual Exploitation Material

“…Literature presents approaches for detecting pornography content or CSEM automatically in images and videos with high accuracy [20,12,14,13,18]. Nevertheless, they require a significant amount of memory and CPU/GPU capabilities.…”