Deep Learning-Based Early Fire Detection Using Small Object Detection Methods

Fires cause severe damage to ecosystems and human life, owing to which developing early detection technologies is critical. Although deep learning models such as You Only Look Once (YOLO) show potential, they cannot easily handle small-scale fire objects owing to information loss during downsampling. This study therefore proposes an enhanced fire detection system based on YOLOv8. Specifically, the proposed method integrates the Content-Aware ReAssembly of FEatures module into the upsampling process to minimize information loss and improve detection performance for small, distant objects. Furthermore, a specialized post-processing algorithm is introduced that utilizes color space rules to eliminate fire-like non-fire objects and a re-detection process for low-confidence detection. Overall, the experimental results demonstrate that the proposed model significantly improves the precision and mean average precision as compared with those under the standard YOLOv8 model. These findings ultimately suggest that the proposed method offers a reliable and practical solution for highly accurate early-stage fire detection in complex environments.