Classification of Physical Violence Actions Using Convolutional Neural Networks with Transfer Learning

Abstract

Violence in the world is a problem of great impact that affects any developing country; it represents a soft system that to date cannot be controlled due to its different manifestations with high rates of crime and delinquency. Artificial Intelligence (AI) uses different innovative resources that help close gaps related to the Sustainable Development Goals (SDGs) such as health, traffic management, climate change, among others, being a way of applying AI to through image processing with Convolutional Neural Networks (CNN). This research evaluates the effectiveness of classifying violent actions such as: strangulation, grappling, kicking or punching, using CNN with Transfer Learning. First, a personalized dataset was created with simulated images of violent actions, made up of 2000 images distributed in 60% for training, 30% for validation and 10% for testing. Second, the pretrained CNN models were trained: VGG16, MobileNetV2, ResNet50 and InceptionV3 applying Transfer Learning, subjected to 150 epochs and using the same hyperparameters. In the end, the performance results were compared between them, where it was determined that the best performance is from MobileNet, which obtained a precision rate of 72.53%, and an accuracy level of 66%. This research will serve as a reference for future applications.