Hybrid CNN-LSTM Model for Automated Violence Detection and Classification in Surveillance Systems

The proliferation of digital images in recent years has prompted the development of automated algorithms that can identify violent and non-violent material in a variety of settings. This study introduces a unique method for categorising violent and nonviolent pictures using a combined Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) model. Careful data collection from a publicly accessible source yielded a dataset of 11,043 original 416x416 photos. The photos were downsampled to 256x256 as part of the preprocessing phase of the suggested technique for more effective model training. The suggested model was compared against both more conventional Machine Learning methods and state-of-the-art Deep Learning systems to determine how well each performed. The findings reported here are for a 30-epoch value and a learning rate of 0.01. The experiment explored several hyperparameters. An outstanding 98.63 % accuracy was achieved by the suggested hybrid model, demonstrating its better performance. This proves the usefulness and promise of using the hybrid CNN-LSTM model for automated image classification, which in turn improves detection and classification skills for violent and nonviolent images.