Lightweight underwater image restoration via frequency-band-aware distillation and cost-aware regularization

Due to wavelength-dependent absorption and scattering, underwater images often suffer from low contrast, colour distortion, and detail loss. Although existing restoration networks achieve strong performance, their high computational cost limits their use in lightweight scenarios. To address this issue, we propose a lightweight underwater image restoration framework based on frequency-band-aware distillation and cost-aware training regularization. The framework uses a high-performance hybrid-domain teacher and a topologically aligned compact student, preserving the teacher’s restoration capability under a reduced parameter budget. To enhance contextual modelling, an omnidirectional cascaded convolution module is introduced at the bottleneck to jointly model two-dimensional contextual degradation and direction-sensitive structures through complementary square, horizontal, and vertical cascaded branches. Frequency-band-aware distillation aligns the spectral magnitudes of intermediate features and outputs, improving texture and fine detail recovery. Meanwhile, cost-aware training regularization incorporates normalized stage-level structural costs into student optimization as a differentiable training-time representation controller, without changing the fixed inference architecture.