For the actual engineering requirements of fast, accurate and real-time detection of ship fire, a lightweight and highprecision SG-YOLO algorithm based on improved YOLOv5s is proposed. The GhostNet convolution structure fused with the parameter- free attention mechanism is used to achieve algorithm lightweight, the 2D attention mechanism and the bi- directional feature pyramid network are introduced to enhance the feature extraction ability of flame, and solve the problems of dense and small flames and imprecise position of target frame. In the self built fire set comparison experiment, compared with YOLOv5s 6.0, the model parameters is reduced by 46.2% , the detection speed is improved by 38.7%, reaching 86 f/s, and the detection accuracy is improved by 0.9%.

ship, flame detection, YOLOv5s, real-time detection