Fire Detection Algorithm for Complex Warehouse Scenarios Based on YOLOv11n

Fire detection in warehouse scenarios is challenged by complex background interference, weak features of small flame targets, and indistinct flame textures, which can reduce detection performance and constrain model deployment. To address these issues, we propose FC-YOLO (Fire Complexity-YOLO), a lightweight fire detection algorithm based on YOLOv11n. A dynamic upsampling module (DySample) is introduced into the neck network to adaptively adjust feature sampling locations and reduce information loss. An adaptive downsampling module (ADown) is employed in the backbone and neck networks to reduce computational complexity while enabling efficient multi-scale feature extraction. A residual multi-scale attention module (IEMA) is constructed in the neck network to enhance feature representation and focus on critical regions. Finally, an additional small-object detection layer (P2) is introduced to improve small-flame detection. On a self-built flame dataset, FC-YOLO achieves 91.8% precision, 84.7% recall, 90.1% mAP@0.5, and 50.0% mAP@0.5:0.95 with 2.2M parameters. Compared with YOLOv11n, these metrics improve by 1.8%, 4.6%, 2.4%, and 3.7%, respectively, while parameters decrease by 14%. These results demonstrate that FC-YOLO improves detection performance and robustness in complex fire scenarios while maintaining a lightweight architecture, indicating strong potential for deployment on edge devices.