EFSD-YOLO: Fire and Smoke Detection Based on Enhanced YOLOv11

Forest fires pose significant threats to human lives, property, and global ecosystems due to their rapid spread and complex terrain, making early detection challenging. Traditional fire detection methods relying on manual patrols or sensors suffer from limitations such as network dependency and terrain constraints, leading to underreporting, missed detections, and delays. With recent advances in image-based object detection, we can now employ broader-coverage, real-time imaging devices for fire monitoring. YOLOv11 (You Only Look Once), which is the latest version in the YOLO series of deep learning object detection algorithms, offers efficient real-time detection capabilities ideal for rapid and precise identification of fire and smoke. However, the basic YOLOv11 framework remains limited when addressing the challenges of wildfire detection, where fire and smoke targets exhibit significant scale variations and complex environmental conditions. To adapt to forest fire scenarios, enhance small object detection and mitigate interference from occlusions, we proposes the EFSD-YOLO based on YOLOv11. First, we enhances the backbone network by incorporating the non-local attention mechanism. This module reduces occlusion interference to improve accuracy through global reasoning. Secondly, a Spatial Attention Multi-scale Fusion (SAMF) module is integrated into the neck network to strengthen feature representation for small targets. Experimental results demonstrate that EFSD-YOLOv11 outperforms YOLOv11n with a $12.3 \%$ higher mAP50 on the CMB-Fire dataset. On the FASDD dataset, it achieves a $2.2 \%$ improvement in mAP50-95, confirming its superior capability for fire and smoke recognition.