Underwater optical imaging exhibits characteristics that are fundamentally different from imaging in air. As light propagates through water, it undergoes wavelength-dependent absorption and scattering, with shorter wavelengths being preserved more effectively than longer ones. This phenomenon causes the well-known blue–green color bias in underwater imagery, accompanied by contrast loss, edge blurring, and severe visual degradation. These distortions negatively affect the robustness of deep learning-based object detectors, whose feature extractors are typically optimized for clear, high-quality terrestrial images. Consequently, underwater object detection remains a significantly more challenging problem than its terrestrial counterpart. In this paper, we propose OceanNet, a multi-scale context interaction framework designed specifically for robust underwater object detection. OceanNet incorporates three key ideas: structural feature preservation, multi-scale contextual reasoning, and task-decoupled detection optimization. A Gradient-Preserved Feature Extraction Backbone is introduced to maintain high-frequency edge information that would otherwise be suppressed by underwater blur. A Cross-Scale Context Interaction Module enables efficient aggregation of global contextual information. Finally, a Dual-Domain Task-Decoupled Detection Head separates classification from localization learning, improving training stability and accuracy. Experiments on URPC2019 demonstrate that OceanNet achieves competitive performance while maintaining computational efficiency[1].