GCE-YOLO: A Lightweight Fire Detection Algorithm Optimized for Edge Devices

In recent years, deep learning based video fire detection methods have been widely adopted in intelligent security systems. However, existing algorithms often exhibit high computational complexity, limiting their feasibility for deployment on edge devices. To overcome this challenge, this paper introduces GCE-YOLO, a lightweight fire detection algorithm based on YOLOv8, designed to enhance efficiency while maintaining detection accuracy. Firstly, the GTFormer module is introduced into the backbone network, integrating the global modeling capability of Transformer with the local feature extraction advantages of CGLU, significantly improving model accuracy and multi-scale fire feature representation. Secondly, a lightweight cross-scale feature fusion module (CCFM) is designed, which reduces model parameters and computational complexity by unifying the channel dimensions in the neck network. Lastly, the EIoU loss function is adopted in place of CIoU for bounding box regression optimization, enhancing the model's adaptability to dynamic changes in fire size by incorporating constraints on center point distance and aspect ratio. Experimental results show that GCE-YOLO achieves an mAP @ 0.5 of 88.6 % and a Precision of 90.6 %, while maintaining a lightweight architecture with only 1.7M Parameters. These findings highlight its efficiency and suitability for fire detection on resource-constrained edge devices.