Fire Detection Based on Surveillance Camera Using Transfer Learning

Abstract

Fire detection is a crucial task within surveillance systems, holding significant importance inpreventing fire-related incidents and protecting individuals and properties. However, thepresence of uncertainties in surveillance, characterized by unclear, incomplete, oruntrustworthy information obtained through surveillance systems, poses challenges to efficientfire detection processes. Additionally, existing fire detection models face several issues,including extensive computational requirements, large model sizes, and many parameters. Theprimary objective of this study is to develop fire detection systems that can accurately detect theexistence of fires in uncertain surveillance environments using transfer learning. To achievethis goal, we utilized four pre-existing convolutional neural network models (VGG19,MobileNetV2, ResNet50, InceptionV3) to extract features specific to fires from images depictingfire incidents. The training phase involved the utilization of a dataset comprising 14,850 images,while the evaluation phase tested the models' performance using 1,650 images. These imageswere gathered from various benchmark research papers and online sources, chosen accordingto specific environmental condition criteria. Hyperparameter optimization further enhanced thesystem's performance. The outcomes of this study revealed a remarkable level of accuracy inthe classification of fire detection, with rates of 98.36% accuracy, 99.27% recall, 98.37% F1-score, and 97.50% precision for VGG19; 95.75% accuracy, 92.37% recall, 95.61% F1-score,and 99.09% precision for MobileNetV2; 98.60% accuracy, 98.06% recall, 98.59% F1-score,and 99.14% precision for ResNet50; and 98.36% accuracy, 99.15% recall, 98.37% F1-score,and 97.61% precision for InceptionV3.These successful results using transfer learning forrobust fire detection in uncertain environments suggest a promising future for improving firesafety through advanced surveillance systems.