Enhanced Fire Detection Using Knowledge Distillation and Multi-scale Attention in YOLOv8

Fire is a destructive disaster that threatens human safety and property. Existing fire detection algorithms often suffer from low accuracy and high omission rates due to varying target scales and complex backgrounds. Additionally, many models are computationally complex, making them unsuitable for real-time deployment on edge devices. To address these challenges, this paper proposes an improved YOLOv8 algorithm based on knowledge distillation, aiming to efficiently compress the model while maintaining detection accuracy. To enhance the model’s ability to focus on key flame features and improve detection accuracy, we introduce a Multi-scale Spatial Attention (MSA) mechanism. To address the challenge of feature fusion in dense fire scenes, we design the Dense Feature Alignment (DFA) module, which refines the fusion process and improves adaptability. Furthermore, to resolve cross-task inconsistencies in knowledge distillation, we propose the Logic Matching (LM) module, ensuring more effective knowledge transfer and consistent feature representations between the teacher and student models. Experimental results demonstrate that the proposed model outperforms existing algorithms in fire detection accuracy. Compared to the baseline model, precision and recall improved by 4.9% and 6.4%, respectively, while mAP@0.5 increased by 3.5%. These results further validate the model’s capability for real-time fire monitoring and accurate fire identification.