Underwater optical imaging suffers from severe degradation due to light absorption, scattering, and suspended particles, leading to low contrast, blurred details, and color distortion, which significantly impairs the accuracy of object detection. To address these challenges, we propose a lightweight yet powerful Dynamic Multi-Scale Perception Network (DMSPNet) for underwater object detection. DMSPNet integrates a Spectral-Guided Feature Extraction Backbone and a Task-Aligned Decomposition Detection Head, enabling coordinated optimization between multi-scale feature representation and detection task alignment. The backbone employs ReefBlock_SFS and AquaMamba_SFS blocks to enhance feature discriminability under degraded conditions. The detection head introduces a task decomposition mechanism with temperature-gated classification and offset-limited regression, improving detection precision for underwater targets. Extensive experiments on the URPC2019 dataset demonstrate that DMSPNet achieves superior performance with 78.2% mAP@0.5 and 50.3% mAP@0.5:0.95, outperforming several baseline detectors while maintaining a compact parameter size of only 2.74M. Ablation studies further validate the effectiveness of each proposed component. The network demonstrates remarkable efficiency and robustness in complex underwater environments, offering a practical solution for real-time underwater perception systems.