A Study on the Performance Analysis of Fire Detection Using a Lightweight Convolutional Neural Network Model

The demand for AI-based fire detection technologies that can swiftly identify various fire and non-fire scenarios is increasing. This study compared the fire detection and classification performances of several CNN-based models across nine fire-related events: clouds, fog, chimney smoke, wildfires, building fires, and indoor fires. Each model was trained using a dataset comprised of 27,000 fire-related images. The experimental results indicated that the MobileNet v1 model, which incorporates a depthwise separable convolution layer structure, was the most lightweight (54.6 MB) while demonstrating moderate classification accuracy (66.0%) in distinguishing between fire and non-fire events. Additionally, a detailed analysis using the confusion matrix revealed a low error rate of 3.78% for misclassifying actual fires as non-fire events, and vice versa. These findings suggest that the proposed MobileNet v1-based fire detection model offers high classification performance for real and non-real fire scenarios, along with the advantages of a lightweight structure, making it well-suited for real-time fire detection in edge computing systems.