Transformer-based extreme selfie detection for enhancing safety in hazardous environments

Abstract

Selfie-taking has become an integral part of modern digital culture, with over 24 billion selfies uploaded globally in 2015 alone. However, the increasing trend of capturing extreme selfies in hazardous environments have led to a significant rise in accidents and fatalities, necessitating the development of technologies for extreme selfie detection to enhance safety by identifying and alerting individuals when they are in potentially dangerous situations. Our paper presents a novel approach to detecting extreme selfies using transformer-based architectures, specifically Vision Transformer (ViT) and Data-Efficient Image Transformer (DeiT). A dataset of 2,000 images, evenly split between 'Normal Selfie' and 'Extreme Selfie' categories, was created from different online portals and social media sources. Extensive preprocessing, including resizing, normalization, and data augmentation, was applied to the training data to prepare the images for classification. The ViT model achieved a superior 98.1% validation accuracy, outperforming traditional models from previous work such as VGG16 and ResNet50, while DeiT performed impressively at 96%, demonstrating its efficiency with smaller datasets. ViT also exhibited high precision, recall, and F1-score, making it highly reliable for detecting dangerous selfies scenarios. The results highlight the potential of transformer models in real-time safety applications, with future work focusing on optimizing these models for mobile deployment and integrating augmented reality to provide live safety alerts during selfie-taking.