LOCDNet: A Lightweight Multi-Target Ship Detection Network for USV Visual Environmental Perception in Dense Occlusion Scenarios

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Reliable visual perception of surrounding vessels is essential for situational awareness and safe navigation of unmanned surface vehicles (USVs). In complex nearshore waterways, densely distributed ships, significant scale variations, and mutual occlusions can degrade onboard monocular visual information, resulting in missed detections and inaccurate target localization. To address these challenges, we propose a lightweight multi-target ship detection network, named LOCDNet, for onboard visual environmental perception of USVs. It is jointly designed from three aspects: boundary preservation, occlusion feature compensation, and lightweight design. A boundary feature enhancement unit is designed for Occlusion-Prior Wavelet Convolution (OPWTConv) to preserve high-frequency contour cues and mitigate edge-feature degradation. Occlusion Recovery Feature Pyramid (ORFP) module recovers incomplete features of occluded ships through multiscale calibration and contextual compensation fusion. Occlusion-Oriented Shared Detection Head (OSDH) module is built upon a shared convolutional structure and incorporates channel attention and occlusion-prior gating, achieving a favorable balance between lightweight and ships’ overlapping detection performance. Extensive experiments on SeaShips and MariShipInsSeg demonstrate that LOCDNet achieves mAP50-95 values of 65.81% and 51.47%, respectively, excelling YOLOv11n by 2.83% and 3.84%. Moreover, it exhibits stronger robustness under various weather and illumination conditions. It can be used for real-time ship perception in complex nearshore scenarios.