Violence Detection Using Deep Learning

Abstract

Detecting violence is important for preserving security and reducing crime against humans, animals, and properties. Deep learning algorithms have shown potential for detecting violent acts. Further, the reach of large and diverse datasets is critical for training and testing these algorithms. In this study, the aim is to detect violence in images using deep learning techniques to enhance safety and security measures in various applications. For that, we adopted the most utilized and accurate models, such as sequential CNN, MobileNetV2, and VGG-16 which are well known in this field to measure the performance for each classification model on a large dataset of annotated images of eight classes containing both violent and non-violent content. The techniques like data augmentation, transfer learning, and fine-tuning are utilized to improve model performance. As a result, the VGG-16 model has achieved a 71% test accuracy that outperform than Sequential CNN and MobileNetV2 with suitable hyperparameters showcasing its potential for integration into surveillance systems, social media monitoring tools, and other security applications.