YOLOv8-enhanced temporal modelling for semantic classification of ethically sensitive video content

This paper presents a smart, efficient, and interpretable deep learning framework for content-based video classification, specifically targeting context-rich categories such as vulgar, kissing, accidental, abusive, and fighting scenes. The proposed approach employs YOLOv8-based object detection to guide an attention-weighted keyframe selection mechanism, thereby ensuring that only the most informative frames are retained for the processing. These selected keyframes were passed through a hybrid deep architecture combining ResNet50 for spatial feature extraction and Bi-LSTM for temporal sequence modelling. Extensive experiments were conducted on benchmark datasets, including UCF101, NudeNet, CADP, and a custom dataset curated for rare categories. An accuracy of 92.34% and an F1-score of 91.12% were achieved by the proposed model, performing better than the traditional models of 3D CNN, Conv-LSTM, and ResNet50 with dense frame sampling by 4-8%. Class AUC values of more than 0.94 were obtained in the analysis of the proposed model.