Abstract: Underwater object detection is a key technology for ocean exploration and development. However, the complexity of the underwater environment and the limited computing resources of underwater equipment pose enormous challenges to underwater object detection. To address these issues, this paper proposes a lightweight underwater object detection algorithm named LSOR-DETR (Lightweight Spatial Online Reparameterization DETR). Firstly, LSPA-Net (Lightweight Spatial Perception Aggregation Network) is proposed by combining the efficient feature fusion mechanism of Cross-Stage Partial Networks with the adaptive spatial modeling capability of weighted convolution, which reduces model complexity while preserving the feature representation ability of the backbone network. Secondly, the ORFusion (OREPA-enhanced Restructured Fusion) module integrating online reparameterization technology and the RepNCSPELAN architecture is presented to construct an efficient feature fusion mechanism. Finally, the GSConvE lightweight downsampling module is introduced to compensate for the precision loss caused by the lightweight process of the backbone network and neck fusion. Experimental results on the DUO underwater object detection dataset demonstrate that the proposed model maintains the detection accuracy of the RT-DETR model while reducing the number of parameters and computational cost by 38.7% and 40.7%, respectively.