<p>High-quality underwater images are crucial for marine research, ecological conservation, and resource development. However, challenges in the underwater environment, such as light attenuation, scattering, and color deviation, significantly degrade image quality. To address these issues, deep learning-based methods have shown great potential in underwater image enhancement due to their ability to learn complex feature representations and adapt to diverse underwater conditions. However, traditional Transformer frameworks often struggle to effectively capture global and local feature relationships in severely degraded and uneven underwater images, leading to suboptimal enhancement results. This study proposes a multi-level fusion framework based on Fourier stabilization and sparse attention mechanisms (FSDSformer) to overcome these limitations. This framework innovatively integrates a Dynamic Sparse Attention Transformer (DSAT), a Fast Fourier Stabilization Model (FFSM), and a Spatial Fusion Attention Module (SFAM), aiming to resolve issues of blurring, distortion, and low contrast in underwater images. FSDSformer enhances image reconstruction by optimizing information transmission within the network, employs dynamic sparse attention mechanisms to refine the fusion of critical information, and leverages Fourier transforms to extract multi-scale features from the spatial domain, ensuring effective enhancement of degraded image regions. Additionally, the spatial fusion attention mechanism optimizes color imbalance between the encoder and decoder. Experimental results demonstrate that the framework performs exceptionally well on multiple public datasets, significantly improving image quality and providing an efficient and reliable solution for underwater image enhancement.</p>