Objective To address the issues of missed detections and false positives in ship detection within coastal environments characterized by dense small targets, severe background interference, and extreme target aspect ratio variations, this paper proposes an improved lightweight ship target detection model named WTDS-YOLO.

Method Based on the YOLOv11n baseline, the WTDS-YOLO framework incorporates three main architectural innovations to optimize performance. First, to overcome inadequate local feature extraction for small targets and enhance noise resistance against sea waves and clouds, we design the C3k2_WT module by integrating lightweight Wavelet Convolutions (WTConv) into the backbone. WTConv utilizes 1-level Haar wavelet decomposition to decouple low-frequency global semantics from high-frequency local details, effectively expanding the receptive field while preserving critical edge features of micro-scale vessels. Second, a Dynamic Multi-Scale Adaptive Feature Modulation (DyMSAFM) module is embedded into the neck network to enhance small target feature representation under dynamic backgrounds. DyMSAFM employs a multi-branch spatial downsampling and channel-split mechanism, combining depthwise convolutions with GELU activation to dynamically generate multi-scale spatial modulation weights that suppress complex background clutter. Finally, the Shape-IoU loss function is employed to optimize bounding box regression for multi-pose and irregular ships. By introducing scale-aware dynamic weighting factors, Shape-IoU enforces strict shape and scale constraints, effectively reducing localization deviation and accelerating model training convergence.

Results Comprehensive experiments were conducted on the benchmark HRSC2016-MS remote sensing dataset (1,680 high-resolution images). The quantitative results demonstrate that WTDS-YOLO achieves an average precision (mAP@0.5) of 87.6%, a recall of 77.8%, a precision of 89.9%, and an F1 score of 83.0%, representing substantial improvements of 5.4%, 5.6%, 4.9%, and 6.0%, respectively, over the original YOLOv11n model. Furthermore, under the stringent mAP@0.5:0.95 evaluation metric, the model reaches 69.5%, reflecting a 7.4% enhancement. Comparative analyses confirm that WTDS-YOLO outperforms mainstream detection algorithms, including Faster R-CNN, YOLOv5n through YOLOv10n, as well as recent YOLOv12n and YOLOv13n baselines. In scenario-specific evaluations, the model achieves 85.0% mAP@0.5 in complex, densely packed port environments and 93.1% mAP@0.5 in offshore scenes. Crucially, while significantly enhancing accuracy, WTDS-YOLO reduces parameters to 2.51 M and computational complexity to 3.08 GFLOPs, achieving an impressive real-time inference speed of 101.56 FPS on an RTX 4070 GPU

Conclusion The proposed algorithm successfully balances high detection accuracy, robust clutter suppression, and computational efficiency, meeting the requirements for rapid, real-time ship detection in remote sensing images.