Real-time obscene scene nudity detection and blurring in a video clip

Abstract

Videos are widely consumed by people of all ages as a form of entertainment, information and education. However, not all videos are made for everyone. Many videos contain obscenities such as nudity, violence, blood, and gore which should not be watched by children or people who feel repulsed by these obscenities. Obscene con-tent can negatively affect a child’s mindset, and it can even traumatize people with weak mental constitutions. The real problem begins when these obscene videos are publicly available on the Internet, and anyone can watch them easily by downloading or streaming them online without getting any kind of warning. Moreover, people can even encounter these obscenities on live video streams or video calls. In our research, we have worked to detect and blur nude and obscene sexual content from videos in real-time. In that respect, this paper proposes a Neural Network-based approach. We have detected whether sexually explicit content is present in a video or not and blurred only the detected contents from the video frames. To detect nude and obscene contents, we have used different object detection algorithms such as Faster R-CNN, YOLOv5 and YOLOv6. These three respectively gave us mean average precision values of 0.382, 0.663 and 0.508 at 0.5 IOU threshold. Although with an mAP value less than YOLOv5, we chose YOLOv6 as it has proved to be the most optimal for our solution in regards of both accuracy and speed. And to blur, we have tried a total of five methods provided by two image processing libraries, OpenCV and PIL. Among those, we have selected the averaging method of OpenCV since it has best suited our needs. Additionally, we have attempted to reduce the rate of false positives so that any decent content does not get incorrectly labelled as obscene. This detection and blurring of obscene contents will contribute to ensuring safety in internet browsing for everyone.