Underwater object detection (UOD) plays a critical role in marine resource exploitation, shipwreck retrieval operations, and underwater ecological monitoring. Despite the advancements in UOD in recent years, this field still presents significant challenges due to the large number of small objects, blurry images, occlusions, and overlaps in underwater scenes. To address these issues, we propose a novel Transformer-based framework, named Re-parameterized Pyramid Detection Transformer (RP-DETR). First, we propose a re-parameterized cross-layer aggregation network (Rep-CLAN), which leverages residual structures to enhance feature representation, thereby improving the detection of small objects. Second, we introduce a hybrid attention module (HAM) to strengthen the distinction between foreground and background information while reducing the impact of image blur. Finally, we design a spatial pyramid average pooling (SPAP) structure for multi-scale feature fusion to effectively detect occluded and overlapping objects. Experimental results on two benchmark datasets, RUOD and URPC2020, demonstrate that RP-DETR achieves superior performance compared to existing state-of-the-art methods in underwater object detection.