Analysis of the Reliability of Fire Detection Using Deep Learning

Automated fire management and suppression systems require reliable and fast fire detection. The present study evaluates the reliability of fire detection based on deep learning. An image classifier has been trained and tested to detect fire using the TensorFlow framework, Keras library, and Python programming language. A sequential model with convolutional layers, max-pooling layers, and dense layers has been built. Classification of geometrical shapes, including rectangles and ellipses, has been used to verify the deep learning model. However, reliable fire detection is far more challenging. Over 30,000 images have been used to train the fire detector. A different set of 3,488 images have been used for testing. The train and test fire datasets include diverse types of fires at different stages of fire growth. Image augmentation and drop-off layers have been used to reduce overfitting. Transfer learning, by inheriting learned knowledge from pre-trained models has been used to improve the accuracy. This combined with multi-stage fine-tuning, by unfreezing groups of top layers in stages, has resulted in a test accuracy of 95.04%. Limitations of the model and the reasons for selected inaccurate classifications have been analyzed. Reliable fire detection needs careful selection of the training images and controlling overfitting. The model presented is capable of reliable fire detection in specific environments, however, developing a global fire detector that can work in diverse environments requires significant progress.