Convolutional neural networks for solving fire detection problems based on aerial photography

The paper presents the results of applying a new structure of convolutional neural networks (CNN) for fire detection based on aerial photographs. A training data set was formed based on aerial video files, taken in various conditions. They show that the proposed convolutional neural network performs quite well in the field of fire detection. The results of experiments on real video sequences are presented. The proposed approach provides high precision 94.78%, recall 92.97%, F1-score 95.42% and IoU (Intersection over Union) value, that shows the effectiveness of the proposed CNN for fire detection.