An Evaluation of State-of-the-Art Object Detectors for Pornography Detection

Pornographic and nudity content detection in videos is gaining importance as Internet grows to become a source for exposure to such content. Recent literature involved pornography recognition using deep learning techniques such as convolutional neural network, object detection models and recurrent neural networks, as well as combinations of these methods. In this paper, the effectiveness of three pretrained object detection models (YOLOv3, EfficientDet-d7x and Faster R-CNN with ResNet50 as backbone) were tested to compare their performance in detecting pornographic contents. Video frames consisting of real humans from the public NPDI dataset were utilised to form four categories of target content (female breast, female lower body, male lower body and nude human) by cropping the specific image regions and augmenting them. Results demonstrated that COCO-pretrained EfficientDet-d7x model achieved the highest overall detection accuracy of 75.61%. Interestingly, human detection of YOLOv3 may be dependent on image quality and/or presence of external body parts that belong only to humans.