Objective Research on ship detection and recognition technologies holds important economic and strategic significance for safeguarding maritime rights and ensuring marine safety. This study focuses on the challenges posed by complex maritime environments, particularly severe weather conditions such as dense fog caused by high seawater vaporization, as well as large-scale glare and irregular water reflections under strong illumination on the sea surface. Furthermore, ship targets often appear very small in large-scale sea−sky scenes, and occlusion frequently occurs in port areas due to the dense distribution of vessels. These factors render traditional space-time background modeling techniques unsuitable for effective marine target detection and recognition. To address these challenges, this research investigates ship target detection and recognition in maritime and aerial scenes.

Method To address the challenges of low accuracy and high false detection rates caused by small targets, densely distributed targets, and various forms of occlusion, this study proposes a novel ship target detection and recognition approach based on an improved YOLOv3 algorithm. The feature extraction network of the original YOLOv3 algorithm is enhanced to strengthen its feature extraction ability. In addition, a fusion strategy combining Res2Net and SE-Net is adopted to enhance the robustness and generalization ability of the algorithm model. The output prediction scales are expanded to more effectively capture small ship targets, thereby improving detection and recognition accuracy. Moreover, by assigning higher weights to small ship targets in the loss function, the false detection rate for such targets is further reduced. Finally, extensive comparative experiments are conducted to evaluate the performance of the proposed method in detecting and recognizing small ship targets, densely distributed vessels, and heavily occluded targets.

Results Experimental results demonstrate that the improved YOLOv3 algorithm achieves accurate detection and recognition of ship targets under the aforementioned scenarios. The proposed method improves the mAP by 9.0% compared with the original YOLOv3 algorithm.

Conclusion The improved YOLOv3 algorithm proposed in this paper effectively addresses the challenges of high omission rates in large–scale maritime scenarios caused by small ship targets, dense distributions, and severe occlusions.