EHD-YOLOv8n: An Enhanced Fire Detection Algorithm with Dynamic Align Detection and Efficient Local Attention Mechanisms

This paper proposes an enhanced fire detection algorithm based on YOLOv8, called EHD-YOLOv8n. The aim of the study is to improve fire detection performance, especially in complex backgrounds and multi-scale targets. The key innovation is the Dynamic Align Detection (DAT) head, designed with a shared convolutional structure that reduces model parameters and computational cost. The DAT head incorporates a scale layer to handle multi-scale targets and a task-aligned mechanism to improve the synergy between classification and localisation. In addition, the Efficient Local Attention (ELA) mechanism is used to improve the HSFPN by integrating multi-scale convolutional kernels and a channel attention mechanism to focus on critical areas. The model was trained and tested on a dataset collected from web pages. Experimental results show that the EHD-YOLOv8n model outperforms the original YOLOv8n model, achieving a mean average precision (mAP) of 84.4% and a detection speed of 183.6 frames per second (FPS). These improvements reduce false positives and negatives while maintaining a reasonable detection speed of 126.0 FPS, ensuring timely fire alerts.