Objectives To address the issues of model redundancy, high computational complexity, and insufficient real-time performance in ship detection within complex maritime environments, we propose a novel vessel detection approach based on the SEE-YOLO model.

Methods Based on the YOLOv12n baseline, we first replace the original backbone with StarNet, which implicitly maps low-dimensional features into a high-dimensional non-linear space via star-shaped element-wise multiplication, substantially reducing the parameter count. Second, we propose a lightweight Efficient-Head detector that further compresses the model architecture through a parameter-sharing mechanism and improves inference speed. Finally, an Efficient Channel Attention (ECA) module is embedded into the Neck structure, enhancing the feature-extraction capability without increasing model complexity.

Results Experiments on the SeaShips7000 dataset demonstrate that the proposed SEE-YOLO model contains 1.27 M parameters and requires 3.3 GFLOPs. Compared with the baseline YOLOv12n, SEE-YOLO reduces the parameter count by 49.2 % and the computational cost by 43.1 %, while achieving 550 FPS, an improvement of 27.3 %. Moreover, the mAP@0.5 reaches 98.8 %. Compared to the latest YOLOv13n, SEE-YOLO achieves comparable performance in terms of mAP@0.5, while reducing the number of parameters by 48%, decreasing computational cost by 47%, and improving FPS by 67%.

Conclusions The SEE-YOLO algorithm achieves a superior balance among detection accuracy, model efficiency, and inference speed, rendering it particularly well-suited for resource-constrained maritime surveillance scenarios.