Enhanced YOLO-security deep learning model for ship detection in maritime security
Zhenbo Bi, Hua Yang, Yingshun Fu, Wenhao Zheng · Results in Engineering · 2025
• We constructed a multi-scene dataset containing 8 types of ships: kayak data is synthesized by virtual reality technology to solve the sample scarcity problem, and low-light ship samples are generated by combining with image synthesis methods, which ultimately results in a comprehensive dataset covering normal/low-light environments. • We propose YOLO-Security, which realizes feature fusion enhancement and ship leakage reduction for robust ship detection in complex lighting scenarios at sea. • We significantly improve the model's precision, recall, [email protected] and [email protected]:.0.95 and also reduce the model's parameter size by integrating ADown, C3_SimAM and SEAM modules in YOLOv5. Due to the complex maritime environment, the variety of ships, and the tendency for ships to be obscured, the accuracy of ship detection in security applications remains low, with a high incidence of missed detections and false positives. Therefore, we propose an enhanced model, YOLO-Security, that can efficiently and robustly detect ships. First, for the insufficient samples of specific types of ships and the missing data of ships in low-light environments, the ship detection dataset is extended by virtual reality technology and data synthesis. Then the Adown (Adaptive Down-sampling Module) downsampling module is introduced to replace the convolutional downsampling layer of the original model Backbone to extract more effective information while reducing the model size; the lightweight attention mechanism is used to enable the model to more accurately localize and identify the target area; finally, the SEAM (Separated and Enhancement Attention Module) attention mechanism is introduced for the ship occlusion problem to enhance the channel correlation and to improve the leakage caused by the occluded target. The experimental results show that YOLO-Security achieves 93.1% and 81.7% for [email protected] and [email protected]:.0.95, respectively, which is better than the original YOLO (You Only Look Once) algorithm in terms of detection accuracy by 13.1% and 25.2% respectively, while the parameters are reduced by 10.4%. This proves that YOLO-Security is more competitive than existing algorithms for maritime security applications.