Ship detection serves as a critical foundation of intelligent maritime systems but remains challenged by complex backgrounds, multi-scale targets, and severe occlusions, all of which substantially compromise detection accuracy. To address these issues, this study proposes an enhanced object detection framework, termed DR-YOLO, built upon the YOLOv8 architecture. Specifically, a Dynamic Receptive Field (DRF) module is introduced to strengthen the backbone network’s capability in capturing multi-scale ship features, while a Recursive Bidirectional Feature Pyramid Network (R-BiFPN) enhances cross-scale feature interaction through recursive fusion. Furthermore, a Task-Cooperative Attention (TCA) mechanism is designed to jointly optimize classification and regression tasks within the detection head. The loss function is further improved into a Dynamic Focal-SIoU formulation, effectively mitigating sample imbalance and improving bounding-box regression precision. Experimental results on the self-constructed ISD-2025 dataset demonstrate that DR-YOLO achieves an mAP@0.5 of 91.6%, surpassing the original YOLOv8 by 1.4%, while maintaining a real-time inference speed of 128 FPS and exhibiting superior robustness under complex maritime conditions.