Enhancing Underwater Imagery: A Fourier-Sparse Attention Transformer Approach

Abstract Underwater image enhancement is a crucial task for marine exploration and computer vision applications. However, existing methods often suffer from severe quality degradations resulting from light absorption and scattering, color distortion, contrast reduction, and detail loss. This paper introduces FST-former, a novel Transformer-based framework that integrates Fourier modelling and sparse attention mechanisms to address these issues. Specifically, the proposed method includes a Multi-Branch Fourier Fusion Module that enhances multi-scale feature extraction capability in both frequency and spatial domains, ensuring effective enhancement of degraded regions. In addition, a Frequency-Adaptive Sparse Attention module is introduced to induce attention sparsity and refine key information fusion, thereby reducing interference from irrelevant regions. Extensive experiments on several public underwater datasets demonstrate that the proposed FST-former achieves superior performance compared with existing advanced approaches in terms of both restoration quality and model efficiency, achieving up to a 21.6576 PSNR and 0.8564 SSIM on the Test-L500 dataset.