Addressing issues such as small-scale targets and complex background interferences that affect the accuracy of ship detection in complex maritime scenarios, this paper proposes RSD-YOLO, an improved lightweight maritime ship detection algorithm based on YOLOv7-tiny. The method begins with the introduction of a lightweight reparameterization module called RepNCSPELAN4, and combines it with a coordinate attention mechanism to improve and obtain the RepELAN_CA structure. Using RepELAN_CA to reconstruct the backbone network enriches feature extraction and enhances target localization accuracy. By decoupling the training and inference stages, the model’s computational load and the number of parameters are significantly reduced. Subsequently, the SPD subsampling strategy, which replaces the traditional method combining striding convolution and pooling. SPD introduced to mitigate the loss of fine-grained features and enhance the detection precision of small-scale targets. Finally, DyHead is introduced to refine the original detection head, improving the model’s capability to handle complex background noise by overlaying two iterations of the DyHead attention blocks. Experimental results on the SeaShips dataset demonstrate that RSD-YOLO increases the mAP by 1.6% to 91.09%, compared to YOLOv7-tiny, and reduces the parameters and computational load by 22.8% and 32.1%, respectively. By achieving high-accuracy ship detection while reducing resource consumption in computation and memory, RSD-YOLO holds significant practical engineering value.