Year: 2026
Type: article
Abstract: Ship detection in complex maritime scenes is challenged by substantial scale variation, limited details of small targets, and interference from sea surfaces and shoreline backgrounds. Preserving target features during feature extraction and fusion while maintaining inference efficiency remains important for maritime monitoring. To address these challenges, this study proposes RSE-YOLOv12, an enhanced ship detection model based on YOLOv12. Receptive field coordinate attention convolution is incorporated into C3k2 to strengthen local feature representation. A Spatial Perception Informed Feature Extraction module replaces A2C2f to combine local features with contextual information. An efficient upsampling convolution block refines upsampled features for cross-scale fusion. Experiments on SeaShips and LEVIR-Ship show that RSE-YOLOv12 achieves [email protected]:0.95 values of 0.805 and 0.316, respectively, exceeding the YOLOv12 baseline by 1.6 and 8.1 percentage points. Inference takes 5.2 ms and 5.4 ms per image on the respective datasets under the experimental settings. These results indicate that the proposed modifications improve ship detection accuracy while maintaining millisecond-level inference in shore-based and remote sensing scenarios.
Source: Scientific Reports
Authors: Jie Wen, Xiaoxia Wan, Yulong Du
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
Award: G3235555129
OpenAlex ID: W7236777260
