Lightweight Fire Detection Network Based on Improved YOLOv8

To address the issues of excessive parameter size and heavy computational burden in existing flame and smoke detection models, which hinder their effective deployment on low-performance devices, this study proposes a lightweight detection network named YOLOv8-GBH based on an improved YOLOv8 architecture. The model aims to achieve an optimal balance between fire detection performance and computational efficiency under limited computing resources. First, a novel lightweight feature extraction module, GhostBottleneckHybrid, is designed. This module combines depthwise separable convolution with residual connections, significantly reducing computational costs while maintaining feature extraction capability. Second, the SimAM parameter-free attention mechanism is introduced, which adaptively reweights feature maps to enhance the model's focus on critical regions. Finally, the feature pyramid network structure is optimized to improve multi-scale feature fusion efficiency. Experimental results on a self-constructed flame and smoke detection dataset demonstrate that the proposed method maintains the original detection accuracy while reducing computational complexity by 50%, decreasing the number of parameters by approximately 68.4%. These improvements significantly enhance the model's deployability on edge devices, providing an effective technical solution for industrial monitoring and other real-time application scenarios.