Dual-domain feature aggregation transformer network for underwater image enhancement

Underwater images often suffer from color distortion, artifacts, and loss of detail due to the refraction and absorption of light in water. These challenges have greatly limited the research in underwater-related fields. However, existing methods only rely on self-attention in the spatial domain to model global information, ignoring potential frequency domain information. To this end, we propose dual-domain feature aggregation transformer network, which improves information capture and utilization through dual-domain feature aggregation to generate detailed and information-rich attention maps. To fully utilize non-redundant information, we propose the frequency-domain enhancement fusion block, which improves model performance by introducing additional enhancement features. In addition, we incorporate hybrid channel upsampling block to further improve the performance and fine textures. Extensive experimental results on commonly used benchmarks demonstrate the good performance of the method compared to state-of-the-art approaches.