Adapt-PEFT: Adaptive Parameter Efficient Fine Tuning for Underwater Image Enhancement

Underwater image enhancement (UIE) is very important for marine exploration and monitoring, yet significant challenges arise from severe image degradation caused by light absorption and scattering in aquatic environments. While deep learning methods achieve promising results, adapting models to new underwater domains through full fine-tuning is computationally expensive and requires extensive training time and domain-specific data. Motivated by the potential of Parameter-Efficient Fine-Tuning (PEFT), we introduce Adapt-PEFT, a novel UIE framework that yields state-of-the-art performance while fine-tuning less than 1% of model parameters. Our key innovation lies in dynamically mixing novel Squeeze-and-Excitation Adapter (SEA) and Scale-and-Shift Features (SSF) components through learnable Gumbel-Softmax routers at each network layer. This adaptive selection exploits the observation that encoder and decoder layers, which capture features at different semantic levels, benefit from layer-specific PEFT strategies. Extensive evaluation on benchmark datasets demonstrates that Adapt-PEFT with 0.172M trainable parameters (0.20% of the full model) performs on par with full fine-tuning. We achieve state-of-the-art SSIM (0.8639) on LSUI-L400 and competitive performance on UIEB-T90, using 78× to 485× fewer parameters than existing methods.