Real-Time Ship Engine Room Initial Fire Identification based on Improved YOLOv8 Network

International trade is facilitated by ships, therefore maintaining ship safety is crucial to the growth of the shipping sector. This study presents a ship engine room fire detection approach based on YOLOv8 algorithm, which aims to address the shortcomings of previous ship engine room fire detection methods, such as low accuracy and inability to reflect the fire state and position in real time. First, the YOLOv8 Backbone is supplemented with the ParNet attention mechanism to improve the model's robustness and capacity to represent features. Subsequently, CARAFE is employed as the upsampling layer to enhance the quality of the feature map and offer the model a more precise feature reconstruction. Lastly, to lessen the effect of poor-quality data on model performance, Focal-EIOU is substituted for CIOU in the model loss function. The enhanced YOLOv8 algorithm has a greater average detection accuracy in ship engine room fire when compared to the SSD, Faster-RCNN, and YOLOv5 algorithms. The [email protected] reaches 97.8%, which is 4.0% higher than the average detection accuracy of the original YOLOv8. A new study plan for the detection of ship engine room fire is provided by the high real-time algorithm's detection speed, which reaches 48.44 FPS.