Treffer: Efficient Fire Detection System Using Custom Deep Learning Models.

Fire detection is a critical task in preventing property damage and saving lives. While many traditional methods rely on sensor data or manual human monitoring, automated fire detection using image analysis has gained significant attention due to the widespread availability of surveillance cameras. However, accurate fire detection system remains challenging aspect. For instance in 2021 alone, forest fires in the United States resulted in property losses of over 23 billion USD, 3,670 human deaths, and 13,350 civilian injuries. Artificial intelligence (AI) and Convolutional Neural Network (CNN) has great potential in many different industries offering superior solutions in several fields. in this paper, Deep Learning (DL) models were investigated (MobileNet, VGG16, and ResNet50) and implemented for fire detection. then we proposed an improved CNN model for early fire detection. The Foggia Fire Dataset was used for both training and testing the models, the models structure include numbers of steps including; pre-processing where the images from the dataset were first scaled, normalised and improved to obtain the best performance, an Augmentation step that includes four types of image transformations: Rotation, Brightness correction, Flipping and Zooming. Dataset Split into three groups training for (80%), validation and testing for (10%). Each model experimented, modified and evaluated using the performance metrics (accuracy, precision, recall and F1-score). an ablation study applied to evaluate the impact of different architectural components on the performance of the proposed CNN models. the proposed CNN model achieved better results than the pre-trained and other baseline models, reaching (99.55%) accuracy, (99.20%) precision, (98.84%) recall, and (98.70%) F1-score, the proposed CNN balances accuracy and efficiency effectively, contributing a practical solution for real-time fire detection applications. [ABSTRACT FROM AUTHOR]