Optimization of YOLO Model Based on Backbone Network Replacement: Fire Detection Performance of LSKNet vs. Benchmark Model

Aiming at the challenges of variable fire scale, complex background interference, and high real-time requirements in the fire detection, this paper proposes, this paper proposes LSKNet-YOLO, an improved YOLO model with a Large Selective Kernel Network backbone. By integrating dynamic large kernel convolution and a feature recalibration mechanism, the model enhances multi-scale flame and smoke feature extraction. On a curated fire dataset (1,579 images, 4,656 labeled instances), the improved model achieves an integrated mAP50 of 0.523 (+7.6% over YOLOv11n), with 8.0% and 11.8% improvements in fire and smoke class accuracy, respectively. Despite a moderate inference speed reduction (14.8 ms/image vs. 0.8 ms/image), the leakage rate in complex scenarios decreases by 10.2%, significantly enhancing early warning reliability in real-world applications such as industrial facilities and densely populated public spaces. This study provides a reproducible framework for balancing accuracy and real-time performance in fire detection, while explicitly addressing deployment challenges through architectural customization.