Abstract
Early fire detection is essential for safeguarding lives and infrastructure. While deep learning-based object detection has shown promise, achieving both high accuracy and real-time performance in fire detection remains challenging. This study proposes an enhanced fire detection approach utilizing the YOLOv9 architecture with optimized fine-tuning techniques. The methodology improves detection accuracy while maintaining real-time processing through systematic model optimization. The key contributions include: (1) development of a comprehensive fire detection dataset covering diverse scenarios, (2) implementation of an effective fine-tuning strategy that preserves essential pretrained features, and (3) optimization of model parameters for fire-specific detection tasks. Transfer learning with carefully tuned hyperparameters is employed to enhance fire event discrimination under various environmental conditions. Experimental results demonstrate that the fine-tuned YOLOv9 model outperforms the baseline in detection accuracy while satisfying real-time processing constraints. The proposed approach offers a practical solution for robust fire detection systems, particularly suitable for deployment in buildings and industrial environments where rapid response is critical.