Learning hybrid dynamic transformers for underwater image super-resolution

Abstract

Underwater image super-resolution is vital for enhancing the clarity and detail of underwater imagery, enabling improved analysis, navigation, and exploration in underwater environments where visual quality is typically degraded due to factors like water turbidity and light attenuation. In this paper, we propose an effective hybrid dynamic Transformer (called HDT-Net) for underwater image super-resolution, leveraging a collaborative exploration of both local and global information aggregation to help image restoration. Firstly, we introduce a dynamic local self-attention to adaptively capture important spatial details in degraded underwater images by employing dynamic weighting. Secondly, considering that visual transformers tend to introduce irrelevant information when modeling the global context, thereby interfering with the reconstruction of high-resolution images, we design a sparse non-local self-attention to more accurately compute self-similarity by setting a top-k threshold. Finally, we integrate these two self-attention mechanisms into the hybrid dynamic transformer module, constituting the primary feature extraction unit of our proposed method. Quantitative and qualitative analyses on benchmark datasets demonstrate that our approach achieves superior performance compared to previous CNN and Transformer models.