<p>Owing to the light absorption and scattering characteristics in underwater environments, underwater images often exhibit color distortion, detail blurring, and low contrast. Current deep learning models are typically trained using a single end-to-end network, failing to incorporate specific prior knowledge of underwater images and leading to inadequately learned network features. Therefore, this study introduces a novel bi-encoder structure by integrating a prior-based encoder into a U-shape Transformer pipeline. Specifically, a multi-scale feature enhancement module is proposed to enrich and combine features at various depths from the two encoders, enhancing the model’s capability to comprehend image details and edges. Compared to the state-of-the-art methods, our approach leverages prior-based features as supplementary information to assist the decoder in better estimating and compensating for degradation induced by water, thereby enhancing object visibility and image detail. Experimental results demonstrate a significant enhancement in both quantitative and qualitative aspects compared to the state-of-the-art methods.</p>