Efficient and lightweight deep learning model for enhanced ship detection in maritime surveillance

Abstract

Accurate and real-time ship detection is crucial for maritime surveillance, enabling effective collision avoidance and route transportation planning. Existing detection methods struggle to balance real-time performance, accuracy, and parameter efficiency, especially in complex maritime environments. To address these challenges, we propose an enhanced lightweight YOLO-based model, named PMS-YOLO, for real-time ship detection. PMS-YOLO introduces a Partial Multi-Scale Feature Extraction (PMFE) module to improve feature extraction capabilities, a Multi-Branch Auxiliary FPN (MAFPN) to enhance feature fusion, and a Shared Convolution Detection Head (SCDH) to reduce computational complexity while maintaining detection precision. Experimental results on self-built and public ship detection datasets demonstrate that PMS-YOLO achieves an average detection accuracy of 81.4 % and 82.0 %, respectively, representing improvements of 1.75 % and 1.74 % over YOLOv8, while reducing the number of parameters by 29.74 % and lowering computational complexity by 12.5 %. The proposed model effectively balances lightweight design and high precision, enabling real-time ship detection and facilitating deployment on devices with limited resources. Our code is available at https://github.com/wsw1996/ship-detection-PMS-YOLO.