Abstract: Airport terminals, sports stadiums, and other large-span spatial venues pose significant challenges for conventional fire detection technologies in meeting early warning requirements due to their unique architectural structures and internal spatial layouts. Existing image-based fire detectors currently have a maximum detection range of 100 meters, while the horizontal spans of many large-span buildings exceed 200 meters, demanding higher detection distances and sensitivity from such detectors. In the early detection of fires in long-distance and large-scale spaces, it is necessary to improve the ability of detection algorithms to accurately identify small target fires. To address this issue, this study proposes an improved YOLOv8-based algorithm specifically optimized for small-target flame detection. This study proposes an improved YOLOv8-based flame image detection algorithm. For enhancing model accuracy, a CA-Res module integrating coordinate attention and dynamic residual adjustment is incorporated. To control model complexity, the BottleneckCSP module in the original model has been optimized. For improving multi-scale detection capability, an additional small-target fire detection layer is added to the Head section of the model's output terminal. Comparative tests demonstrate that the three-dimensional improvements to YOLOv8 significantly enhance flame detection accuracy while maintaining satisfactory detection precision and real-time performance for small-target flame images. This technical solution effectively addresses the challenges of long-distance and wide-range fire detection in image-based fire monitoring systems, providing a viable approach for early fire warning in large-span spaces such as airport terminals and stadiums.

Key words: fire detection, target detection, deep learning, small target;flame image detection