Underwater biological target detection still predominantly relies on manual identification methods, facing challenges related to low levels of intelligence. Existing target detection algorithms, such as the YOLO series, suffer from issues such as large parameter counts, high computational requirements, and poor detection accuracy. This paper proposes an improved algorithm based on the RT-DETR model. The DynaShareNet backbone network is introduced, which shares stem information architecture to enhance feature fusion efficiency and reduce computational burden; the Dilated Transformer Attention Block (DTAB) is introduced to combine global and local feature interactions to enhance robustness in complex underwater environments; the MaSA-RetBlock module is adopted to address target blurring and low-contrast recognition issues; and the EMASlideVarifocalLoss is introduced to enhance the ability to handle difficult-to-classify targets. Experimental results on the URPC2020 dataset demonstrate that the improved algorithm significantly enhances detection accuracy, with mAP50 and mAP50:95 improving by 3.3% and 3.5%, respectively, while significantly reducing model complexity, with parameter counts and computational costs decreasing by 41.7% and 47.7%, respectively. The detection accuracy and parameter count/computational complexity outperform YOLO series algorithms, and the algorithm demonstrates excellent generalization performance on the RUOD dataset. The study indicates that the improved algorithm effectively enhances the performance and efficiency of underwater target detection, offering promising application prospects.