Spectral-Semantic Decoupled Rectification Network for Non-Uniform Underwater Image Restoration

Underwater image restoration is severely hindered by a tightly coupled degradation process: wavelength-dependent spectral distortion combined with non-uniform, multi-scale spatial scattering. Standard Convolutional Neural Networks (CNNs) and rigid physical priors frequently fail in these dynamic environments, limited by restricted receptive fields, overlooked inter-channel spectral correlations, and severe over-enhancement in photon-starved regions. To break this bottleneck, we propose the Phased Feature Rectification Network (PFR-Net), a decoupled architecture that transforms the ill-posed restoration task into a sequential global spectral calibration and deep semantic refinement paradigm. In the first phase, an efficient Multi-Layer Perceptron (MLP)-based Color Mapping (MLP-CM) module acts as a front-end calibrator to rectify global color casts with minimal parameter overhead. Following this, semantic structures are systematically reconstructed through Dilated Multi-Scale Residual Blocks (DMSRB), which are designed to prevent the truncation of delicate structural gradients and dynamically stabilize content-adaptive feature propagation. To bridge the gap between spatial processing and spectral consistency, we introduce Cross-Covariance Global Attention (XCA), capturing long-range inter-channel dependencies with a linear spatial complexity of