Fire Classification and Early Detection Using Convolutional Neural Networks Based on MobileNetV2 with Transfer Learning

25 12

Özet

Since fires pose a great threat to human life, ecology, and economy, their early detection is a very important issue. In this research, a convolutional neural network model based on the MobileNetV2 architecture was developed and tested using the transfer learning method for automatic fire classification. The study used a binary database created as part of the NASA Space Apps Challenge 2018 competition, which consists of 755 outdoor images with and without fires. Despite the uneven nature of the database, an equal distribution of data was carried out in the pre-processing and validation set, taking into account the class imbalance. The model was initialized with MobileNetV2 weights pre-trained on the ImageNet database, and the final layers were adapted by fine-tuning. The training process was carried out with the Adam optimizer, categorical cross-entropy loss function, and early stopping callback. As a result, the model achieved 98.00% accuracy on the test set, and the test cost was 0.07776. According to the classification report, the precision value of the fire class was 0.98, the recall value was 1.00, and the F1-score was 0.99; the no-fire class was 1.00, 0.90, and 0.95, respectively. Confusion matrix analysis showed that all fire images were correctly identified and that there were only a few false positives. The simplicity of the MobileNetV2 model allows it to be used on resource-limited devices. The results obtained exceed those in the literature. The study contributes to the improvement of early fire detection systems and provides a ready-to-use solution for real-time monitoring.