Indoor Fire and Smoke Detection Using Soft-Voting Based Deep Ensemble Model

Fire and smoke detection using vision-based technology plays a crucial role in terms of safety for indoor environments. In the literature, there are several deep learning-based fire detection solutions available, but most of the existing solutions suffer from low accuracy, high false alarm rates, and vanishing gradient issues. To overcome these issues, this paper proposed a soft-voting based deep ensemble model for fire and smoke detection tasks in which four transfer learning models such as MobileNetV2, ResNet50V2, EfficientNetB0, and DenseNet121 are used as base learners. The proposed model has a 99.11% accuracy rate, a 97% precision rate, a 98% recall rate, and a 98% F1-score.