Advancing fire detection: two-stage deep learning with hybrid feature extraction using faster R-CNN approach

Fire incidents pose severe threats to life, property, and the environment, accounting for significant losses worldwide. Traditional sensing technologies exhibit limitations in effectively detecting fires, particularly in larger spaces. The application of deep learning techniques on fire detection systems has been widely explored. However, many challenges are associated with fire detection technologies, especially in scenarios like indoor fires and forest fires, as well as whether the fire is accompanied by smoke or not. Which results in substantial environmental losses and long-term recovery periods. Early sensing technologies lacked effectiveness in detecting fire instances in open spaces due to response delays and failure to utilize static and dynamic features. In this paper, we aimed to address these challenges by proposing a two-stage fire detection approach using deep learning techniques. The approach proposed a new Faster R-CNN architecture, including our proposed hybrid feature extractor. Evaluation yields a mAP@0.5 of 90.1% with an accuracy of 96.5%. The outcomes indicate that our new hybrid feature extractor surpasses the effectiveness of conventional single backbone transfer learning methods and Yolo’s one-stage detection approach in accurately identifying flames and smoke in various indoor and outdoor settings, providing an accurate fire detection system.