Improving the Performance of Fire Detectors Using Generative AI

Fire incidents are difficult to reproduce under controlled conditions, which leads to substantial limitations in obtaining sufficient image data for training deep-learning-based fire detection models. Because image-based fire recognition systems depend highly on the diversity and volume of training data, effective data augmentation strategies are necessary to overcome such data scarcity. In this study, a method was developed to enhance fire detection accuracy and sensitivity by generating synthetic fire images using generative artificial intelligence (generative AI) and incorporating them into the training dataset. The proposed data augmentation method integrates conventional image transformation techniques, such as rotation and flipping, along with a CycleGAN-based fire image generation approach that can be applied to unpaired image sets. Experimental results showed that augmenting the training data using CycleGAN significantly improves the fire detection performance compared with models trained solely on real-world images.