Cross-Domain Underwater Image Enhancement Guided by No-Reference Image Quality Assessment: A Transfer Learning Approach

Abstract

Single underwater image enhancement (UIE) is a challenging ill-posed problem, but its development is hindered by two major issues: (1) The labels in underwater reference datasets are pseudo labels, relying on these pseudo ground truths in supervised learning leads to domain discrepancy. (2) Underwater reference datasets are scarce, making training on such small datasets prone to overfitting and distribution shift. To address these challenges, we propose Trans-UIE, a transfer learning-based UIE model that captures the fundamental paradigms of UIE through pretraining and utilizes a dataset composed of both reference and non-reference datasets for fine-tuning. However, fine-tuning the model using only reconstruction loss may introduce confirmation bias. To mitigate this, our method leverages no-reference image quality assessment (NR-IQA) metrics from above-water scenes to guide the transfer learning process across domains while generating enhanced images with the style of the above-water image domain. Additionally, to reduce the risk of overfitting during the pretraining stage, we introduce Pearson correlation loss. Experimental results on both full-reference and no-reference underwater benchmark datasets demonstrate that Trans-UIE significantly outperforms state-of-the-art methods.

Figures (8)

• Figure 1: Examples from LSUI article14 and UIEB article12 benchmarks. The pseudo labels shown in (b) exhibit issues such as color cast and over-enhancement.
• Figure 2: Illustration of our framework Trans-UIE. Trans-UIE is built upon a transfer learning framework that consists of both pre-training and fine-tuning stages. During pre-training, Trans-UIE learns from reference datasets. To prevent confirmation bias, we incorporate an above-water NR-IQA model with frozen parameters to guide the fine-tuning process. Additionally, a dataset composed of both reference and non-reference datasets is utilized for fine-tuning to mitigate overfitting, reduce distribution shift, and bridge the gap between pseudo labels and the above-water image domain.
• Figure 3: An overview of the proposed Pre-Train Network (UIR-Net).
• Figure 4: Visual comparisons of full-reference data from LSUI-400 article14 and UIEB-90 article12 benchmarks. Besides, the ground truths are displayed in the last columns.
• Figure 5: Visual comparisons on non-reference benchmarks UIEB-60 article12, EUVP article44 and RUIE article65.