Small Fire Detection Based on Improved YOLOv5

Fire detection is crucial for maintaining social order and ensuring the proper functioning of societal systems. Object detection is a key aspect of edge computing, yet the complexity of detection models and the challenges in identifying small fire pose significant difficulties. This paper presents a lightweight YOLO model for small fire detection, named R-YOLO. The GhostNet module replaces certain convolutions in the YOLOv5 model to create a more efficient architecture. Additionally, a Coordinate Attention mechanism (CA) enhances localization accuracy for small targets, while the Spatial Pyramid Pooling layer is replaced by an ASPP module with parallel dilated convolutions, improving detection precision. Experimental results show that R-YOLO achieves a 2.8% increase in precision, a 6.7% boost in recall, and a 7.3% rise in [email protected] compared to YOLOv5. Deployed on Raspberry Pi 3B+, JetsonNano, MLU370, Orange Pi, XM-RK3588, and Huawei Atlas 500, the R-YOLO implementation achieves 59.5 FPS and 88.7% accuracy on the MLU370, demonstrating excellent performance and highlighting its potential for small fire recognition. demonstrating its potential for small fire recognition.