Underwater image superresolution technology faces the dual challenges of complex degradation patterns and the demand for real-time processing on mobile devices. Existing works have preliminarily addressed the coupled modelling of underwater degradation patterns (e.g., color shifts, scattering, and blurring) through multibranch network architectures. However, most models suffer from high parameter counts and computational complexity, making them impractical for real-time mobile deployment. This paper proposes a lightweight dual-path fusion network that achieves high-precision real-time underwater image reconstruction through a physics-prior-guided feature correction module and a dual-path collaborative feature extraction mechanism. The network adopts a three-stage architecture design: the preprocessing module generates low-resolution samples and geometric encodings via coordinate parameterization; the dual-path feature extraction module combines global context modelling with multiscale local detail enhancement; and the reconstruction module leverages frequency-domain basis functions to guide efficient decoding and generate reconstructed images. Experiments demonstrate that the proposed model achieves significant PSNR improvements over the mainstream RDLN model in × 2 superresolution tasks across different datasets, with a nearly 30% reduction in parameter count.