Underwater object detection faces critical challenges including degraded optical conditions, inter-object occlusions, and class imbalance in marine datasets and environments. This paper presents CDS-YOLO, an enhanced architecture based on YOLOv10s, with three key innovations. First, we devise the C2f-Omnidimensional Dynamic (C2f-OD) module incorporating omnidimensional dynamic convolution to adaptively enhance feature representation for blurred and occluded targets. Second, an Efficient RepGFPN with Depthwise Separable Convolutions (DSERepGFPN) is proposed, which integrates structural re-parameterization with depthwise separable operations to achieve a large reduction in computational load while maintaining multi-scale feature fusion capacity. Third, a hybrid SlideLoss-Focaler-WIoU (SFW) Loss Function is developed to address class imbalance through dynamic sample weighting and geometry-aware regression optimization. In terms of model evaluation, ablation studies confirm each component contributes 0.3–1.1% mAP improvements. Comprehensive evaluations on the RUPC dataset (13,680 images, 10 object categories) demonstrate that CDS-YOLO achieves 87% mAP50, outperforming YOLOv10s by 1.9% and YOLOv12s by 2.3% with comparable computational cost. This work provides a functional solution for efficient underwater detection in resource-limited scenarios.