Automated censoring of cigarettes and liquor drinking in videos using deep learning techniques

Abstract

Exposure to content depicting smoking, alcohol consumption, and other addictive behaviors on social media platforms has been linked to an increase in youth engagement with these substances. This trend is concerning, as early exposure to such content can normalize substance use and lead to initiation among impressionable youth. Therefore, automatic censoring of such content is essential to ensure alignment with community standards and legal regulations, protecting users from exposure to inappropriate material. This research introduces A-Censor, an advanced deep learning system designed to automatically detect and censor smoking and alcohol consumption in videos, addressing critical public health concerns. The development process encompassed data collection, model training, and evaluation. In the data collection phase, a dataset of 3,000 images was assembled, comprising neutral, alcohol consumption, and smoking instances. Feature extraction was performed using MobileNetV2, while classification was conducted using algorithms such as Faster R-CNN, RNN, Gradient Boost, and SVM. The optimal model, Faster R-CNN with a MobileNet backbone, achieved a superior accuracy of 93.53%, outperforming other models. Following detection of smoking and alcohol consumption instances, a blurring technique is applied to obscure harmful content while preserving video quality. A-Censor offers an efficient, automated solution for content moderation, promoting a healthier digital environment.