A lightweight visual detection method for maritime autonomous surface ships in port area navigation

Abstract

Accurate and real-time ship detection in complex port environments is critical for the safe navigation of intelligent ships. Compared to open waters, port areas feature narrow waterways, dense obstacles, and variable lighting, which impose stricter requirements on detection accuracy. Existing one stage detection models, while efficient, often suffer from excessive parameter size, high computational complexity, and insufficient optimization for port-specific challenges. Moreover, port ship image data is scarce, and traditional data augmentation methods are inadequate for generating effective training samples, resulting in poor model generalization. To address these issues, this paper proposes a lightweight visual detection model combining EAE-DCGAN and EAE-YOLO. By introducing Triplet Attention into DCGAN, EAE-DCGAN is proposed. It generates diverse port ship images to enrich the ship dataset. By integrating the LDHS Head, the Triplet Attention mechanism, and the Focal EIoU loss function, EAE-YOLO is proposed, which reduces model parameters and computational complexity while ensuring detection accuracy. Experimental results demonstrate that the proposed method achieves improved detection performance compared to YOLOv10n. Meanwhile, it reduces parameters by 21.74 %, FLOPs by 16.42 %, and model size by 25.58 %, while increasing FPS by 9.66 %. Real ship target detection results further validate the superiority of the proposed method.