Analysis of Fire Prediction Performance of Image Classification Models based on Convolutional Neural Network

Abstract

In this study, fire prediction performance was analyzed using convolutional neural network (CNN)-based classification models such as MobileNetV2, ResNet101, and EfficientNetB0 applicable to an edge computing-based fire detection system for improving fire safety. The fire prediction performance was evaluated using the performance evaluation measures including accuracy, recall, precision, F1-score, and the confusion matrix. The model size and inference time were assessed in terms of the light-weight classification model for the practical deployment and use. The analysis results confirmed that the EfficientNetB0 model had the highest fire prediction accuracy, and the MobileNetV2 was the best light-weight classification model. Notably, additionally learning the image features about light and haze images having similar features with those of the fire images improved the fire prediction accuracy of the light-weight MobileNetV2 model.