Evaluation of Convolution and Attention Networks for Nudity and Pornography Detection in Sketch Images

Internet has become a main source of information for people in numerous areas. However, internet can be useful or harmful according to types of contents exposed. When the visual sexual and violent contents are available, the danger targets children and youths and causes a damage in their mental health. Consequently, content moderators are required to be employed in all websites and social media platforms. Several companies involve human to review, filter, and edit visual sexual and violent contents. However, manual reviewing and filtering is costly and tough. Therefore, big companies such as Amazon and Microsoft used machine learning based solutions to automate the content moderation process. This paper proposes a new solution for content moderation that targets sketch images specifically. Three state of the art models including EfficientNetB7 convolutional neural network (CNN), vision transformer, and convolution and attention (CoAtNet) have been employed for this task to detect nudity and pornography in sketch images. A novel benchmarking dataset including 4417 sketch images were used for training and testing. It was found that our proposed methods outperformed the cloud-based content moderation solutions such as AWS, and Microsoft Azure in terms of accuracy (97%) and F1-score (97%).