DMC-YOLO: An improved YOLOv12-based approach for ship detection in distributed acoustic sensing images

Ship detection plays a vital role in ensuring maritime transportation and navigation safety. However, achieving accurate and efficient detection in distributed acoustic sensing (DAS) images remains challenging due to unique linear texture features, complex hydroacoustic background noise, and significant signal scale variations caused by vessel speed differences. To address these challenges, this study proposes DMC-YOLO, built upon the lightweight YOLOv12n baseline, introducing three key innovations: C3k2-DMSF, Multi-Scale Selective Triplet Attention (MSTA), and Content-Enhanced Dynamic Head (CEDH). C3k2-DMSF replaces the standard bottleneck layer, employing multi-scale receptive fields and dynamic feature fusion to capture multi-scale ship features caused by speed variations. MSTA utilizes selective kernels and triple gating to enhance linear wake textures while suppressing background noise. CEDH preserves critical details under complex sea conditions through content enhancement and a dynamic detection head, improving localization and classification robustness. Experiments on the DAShip dataset demonstrate that, compared to YOLOv12n, DMC-YOLO achieves a 3.51% relative improvement in [email protected] for ship detection. For the four dominant vessel categories in the DAShip dataset, DMC-YOLO attains an average precision of 89.6% and [email protected] of 87.7%, representing relative improvements of 6.5% and 5.3% over the baseline, respectively. In conclusion, DMC-YOLO provides the first dedicated, real-time, and high-performance solution for DAS-based ship detection.