Lightweight Forest Flame Smoke Detection Algorithm Based on Yolov5

Abstract

Forest smoke and flame detection is dominant in ensuring forest safety. To extinguish fire sources promptly and prevent the spread of wildfires, this paper proposes a lightweight improved YOLOv5 algorithm that is more accessible to embedded devices. The proposed algorithm is based on two pivotal ideas: (1) Introducing a normalization-based NAM attention mechanism into the neck network, which suppresses insignificant weights through weight sparsity penalties to enhance key feature extraction. (2) Incorporating a Slim-neck structure, which leverages GSConv and VoVGSCSP to construct a lightweight neck network. Research findings indicate that the optimized network outperforms the baseline YOLOv5s architecture, with a 3.34% gain in precision (P), while reducing FLOPs by 8.8%. This ensures a more lightweight model while maintaining detection accuracy.