MDP-YOLO: underwater object detection based on enhanced feature extraction

Underwater object detection is crucial for ocean exploration and conservation but remains challenging because of poor illumination, a high proportion of small objects, weak textures, and large-scale variations in underwater images. To address these issues, we propose MDP-YOLO, an underwater detector built on the YOLO11 baseline. The method focuses on feature-extraction enhancement and detail-preserving upsampling. In the backbone, we design an MDRFF module to jointly aggregate directional spatial statistics, frequency cues, and global context. In the neck, we replace fixed interpolation with DySample, which predicts content-adaptive sampling offsets and uses bilinear grid sampling to preserve high-resolution details. In the detection head, we add a P2 branch and prune P5 to better match the small-object scale distribution. For box regression, MPDIoU replaces CIoU to strengthen corner-level geometric alignment and reduce boundary jitter. Experiments show that MDP-YOLO achieves mAP 0.5 values of 85.3%, 84.9%, and 86.7% on URPC2020, URPC2021, and RUOD, outperforming YOLO11n by 2.9, 2.2, and 2.5% points, respectively. Overall, MDP-YOLO consistently improves detection accuracy while retaining a compact parameter count across different underwater datasets.