BLN-YOLO: A Lightweight Neural Network Framework for Enhanced Forest Fire Detection

摘要

This paper introduces BLN-YOLO, a lightweight model improved from YOLOv8n, addressing two key challenges in forest fire detection: poor real-time performance of traditional methods and the accuracy-efficiency trade-off in lightweight architectures. The model employs a three-stage optimization: 1) Integrating Local Spatial Attention (LSKA) in the backbone to enhance fire feature extraction; 2) Adopting a cross-scale BiFPN feature pyramid in the neck for efficient multi-scale fusion and small-target detection; 3) Replacing detection heads with Low-rank Decomposed (LSCD) structures to reduce parameters. For edge deployment, a two-stage compression is proposed: structured pruning via LAMP criteria removes redundant weights, while knowledge distillation transfers teacher model's feature responses and decision boundaries to mitigate accuracy loss. Experimental results show BLN-YOLO achieves 79.4% parameter reduction, 0.6% mAP improvement, 77.4% memory savings, and 11.4% faster inference compared to YOLOv8n, meeting ultra-real-time detection requirements.