Single underwater image enhancement (UIE) is a challenging ill-posed problem, and its development is hindered by two main issues: (1) The labels in existing underwater reference datasets are pseudo ground truths with severe color distortions; relying on them in supervised learning leads to domain discrepancy. (2) Underwater reference datasets are limited in size, and training on such small datasets tends to cause overfitting and distribution shifts from real underwater conditions. To address these problems, we propose a transfer learning-based UIE method—Trans-UIE. This method captures the fundamental paradigms of UIE through pretraining and utilizes a hybrid dataset for fine-tuning, narrowing the distribution gap from real underwater environments. We introduce the most suitable no-reference image quality assessment (NR-IQA) metrics from above-water scenes into the cross-domain transfer learning process to correct the distortions caused by pseudo labels, reduce the domain gap between underwater and above-water images, and avoid confirmation bias. Additionally, we introduced a Pearson correlation loss function to mitigate overfitting during the pre-training phase. We employed a Channel Reordering Gated Feed-Forward Network (CRGFN) to enhance the network’s capability in modeling complex degradation factors. Experimental results on both full-reference and no-reference underwater image benchmark datasets demonstrate that Trans-UIE achieves outstanding performance in image enhancement quality.