Conventional fire detection systems, which depend on smoke and fire sensors, frequently have problems with calibration, susceptibility to environmental changes, and the requirement for ongoing maintenance. These flaws can result in unfavorable false alarms. Reducing the amount of damage caused by fire requires prompt and precise identification of potentially dangerous fire incidents. In response, this paper proposes a vision-based fire and smoke detection system based on computer vision and convolutional neural networks (CNNs). The main goal is to create a novel model that is built on three different architectures: Inception-v3, VGG-19, and VGG-16, each of which is customized by adjusting layers and hyperparameters. For a thorough assessment, the research integrates two publicly available datasets, BoWfire and Fire-Flame, and presents a brand-new NoFiSe dataset. Results show that the Inception-v3 model is quite effective; it outperforms other techniques with 91.11%, 94.99%, and 98.03% accuracy rates in the Fire-Flame, NoFiSe, and BoWfire datasets, respectively. The main novelty of this work is the integration of customized multi-scale CNN architectures with a newly developed multi-class fire/smoke dataset to achieve robust and generalized early fire detection under diverse real-world conditions. The study highlights the potential of intelligent fire suppression and emphasizes the need for continued research to improve current models in order to reduce future fatalities.
Janibeigi et al. (Fri,) studied this question.