ADV-YOLO: improved SAR ship detection model based on YOLOv8

Synthetic aperture radar (SAR) ship detection plays a crucial role in supporting maritime traffic control, sea rescue, and marine environmental protection. Despite its importance, SAR ship detection confronts several challenges, including the small size of ship targets, unclear contours, complex background noise, and variable scales of ships. To address these challenges, this paper introduces an enhanced SAR ship detection model, termed ADV-YOLO, which builds upon the YOLOv8 framework. The proposed model incorporates space-to-depth building blocks to improve detection accuracy for low-resolution images and small objects. Additionally, a dilation-wise residual module replaces the C2f module in the network’s neck, augmenting the model’s capability to discern multi-scale targets and enrich feature representation. Furthermore, the WIoU loss function is adopted to replace the conventional CIoU loss, enhancing model accuracy, particularly for low-quality sample bounding boxes. Extensive experiments conducted on the HRSID and SSDD datasets demonstrate the robustness and reliability of ADV-YOLO. Compared to YOLOv8n, there is a significant performance improvement: the proposed method achieves an AP_50-95 of 70% on the HRSID dataset, with an improvement of 4.5%. Additionally, it improves by 3.1% for AP₅₀ and 5.7% for AP₇₅. On the SSDD dataset, the AP_50-75 improves by 0.9%, AP₅₀ by 1.1%, and AP₇₅ by 0.9%. This advancement underscores the potential of ADV-YOLO in enhancing real-time maritime surveillance and safety applications.