Real-Time Harmful Content Recognition in Videos

With the boom of OTT platforms, movie viewing has increased significantly. To enhance the overall family viewing experience and to ensure that the content being viewed is suitable for all age groups, the content needs to be flagged as appropriate, violent, and adult categories. Real-time video content analysis can classify video segments as they are being streamed. This paper investigates several deep learning models, including 3DCNN, slow-fast, slow-fast with ResNet50, and VideoMAE, for real-time harmful content recognition in video. Based on the experimental results, the proposed VideoMAE, a video classification model fine-tuned on the VidHarm dataset, resulted in an accuracy of 67%. This approach not only overcomes language barriers by providing a language-independent model for recognizing harmful video content but also offers language-independent captions for portions identified as inappropriate, whether due to violence or adult content.