SVN-YOLO: a high-precision ship detection algorithm based on improved YOLOv10n SVN-YOLO: a high-precision ship detection algorithm

• Published in: The Journal of supercomputing Vol. 81; no. 14
• Main Authors: Wang, Zixuan, Han, Dezhi, Han, Bing, Wu, Zhongdai, Huang, Xiaohu
• Format: Journal Article
• Language: English
• Published: New York Springer US 31.08.2025
• Subjects: Compilers; Computer Science; Interpreters; Processor Architectures; Programming Languages; NWDLoss; Ship detection; YOLOv10n; Feature fusion; Small targets
• ISSN: 1573-0484
• DOI: 10.1007/s11227-025-07743-4

Ship detection and identification are essential for effective maritime traffic management and the assurance of navigational safety. In light of the challenges posed by the presence of large targets, severe dense occlusion interference, and a considerable omission rate of small targets in marine scenes, this paper introduces a novel ship detection algorithm that builds upon the improved YOLOv10n, referred to as SVN-YOLO. The algorithm features three innovative optimizations compared to the benchmark model. First, the Spatial Pyramid Decomposition Convolution (SPDConv) module is introduced to reconstruct the feature pyramid network, enhancing multi-scale feature representation capability through spatial information compression and channel dimension extension. Next, the VoVNet and GSConv Cross-Stage Partial Connection (VOVGSCSP) with a variable receptive field feature fusion structure is employed to replace the traditional C2f module, utilizing a gating mechanism to enhance feature reuse efficiency and reduce computational redundancy. Furthermore, the boundary regression process is optimized using the Normalized Wasserstein Distance Loss (NWDLoss), thereby improving localization accuracy for dense targets. Experimental results from the SeaShips and InfiRay ship datasets show that SVN-YOLO significantly improves detection accuracy while maintaining real-time detection efficiency. Key performance metrics substantiate these advancements: mAP@0.5 achieves 98.9 %, while mAP@0.5:0.95 increases by 1.3–0.2%. An increase of 1% is observed in Precision, and Recall improves by 0.4%. Meanwhile, Params and FLOPs are reduced by 4% and 10.7%. The proposed algorithm offers a reliable technical solution for intelligent ship detection in complex marine environments.