Image-based deep learning architecture for indoor fire and smoke detection

Abstract

Fire incidents are becoming more frequent in both indoor and outdoor settings. To ensure the safety and security of indoor property, it is essential to have an effective fire detection method. Traditional detection techniques often depend on smoke, heat, or fire sensors. Deep learning models have outperformed traditional color features-based fire and smoke detectors by reducing false alarm rates and extending the detection range. Most deep learning models approach fire and smoke recognition as a classification task and neglect the region proposals. Deep learning-based object detectors have superseded them by incorporating these proposals and providing significant accuracy as well as lower false positive rates. However, existing object detectors fall short of improved accuracy in indoor settings and real-time detection, considering fire and smoke dynamics. This study aims to overcome these challenges by developing FireNet-D5 model for effective fire and smoke detection in indoor settings. The FireNet-D5 model integrates a Squeeze and Excitation (SE) attention module at the backbone and a multi-feature fusion BiFPN module at the neck and adds color filtering with NMS to the pre-processing stage. Later, the model was optimized by a genetic algorithm to determine the optimal hyperparameter settings. To cover all prospective indoor scenarios, a novel medium-scale indoor fire and smoke (ISFire) dataset is developed. It consists of four classes (i.e., Blue Fire, Black Smoke, Yellow-Orange Fire, and White-Gray Smoke), with manual annotation. There are 7,322 images in the dataset. An explainable AI method, Grad-CAM, is used to provide visual explanations of model predictions, ensuring interpretability and transparency. Finally, for a fair evaluation, seven benchmark object detection models are re-implemented, along with the proposed model (FireNet-D5), and trained them on the ISFire dataset. In addition, a statistical metric, "Cohen's d", is applied to test the practical significance and consistency of the models' outcome. The experimental results indicated that FireNet-D5 improved the baseline network's mAP@0.5 by 3.0%, reduced the model parameters by 2.7%, and showed a 32% comparative difference based on Cohen's d effect size. Ultimately, this research lays the foundation for future developments in reliable indoor fire detection technology.