Improving CNN-Based Video Violence Detection
Iulian Vlascu, Alexandra Băicoianu, Mihai Ivanovici · 2025
This study evaluates two convolutional neural network (CNN) architectures designed for video-based violence recognition: a compact, enhanced 3D CNN inspired by DenseNet, optimized for efficient and robust spatiotemporal feature extraction, and a deeper, more computationally intensive 3D CNN, used as a baseline for comparative analysis. Furthermore, an ablation study is conducted to systematically investigate the individual contributions of specific architectural components to the overall model performance. In addition, class activation mapping is integrated into the methodology to facilitate the development of an intelligent cropping mechanism for data preprocessing. We conclude that a reduction in model parameters can be achieved with no impact on performance. By integrating an intelligent cropping mechanism, substantial improvements were observed, exceeding 80% across multiple evaluation metrics, even in the case of smaller model configurations.