MISU: Multi-Input Swin-UNet

Underwater image enhancement (UIE) plays a crucial role in underwater imaging for ocean exploration. However, a reliable UIE method remains elusive, with current approaches either facing limitations in generalizability and dynamic adaptability or neglecting the differences in receptive fields and the depth of semantic information across scales. To address this, we propose multiinput Swin-UNet (MISU), a highly effective and robust UIE framework. MISU incorporates the concept of coarse-to-fine into the UIE task. Specifically, to effectively manage the diverse resolutions from multiple scales, we design a multiple-resolution input strategy. Then, a Swin-Transformer-based block is introduced to extract the feature from multiscale, thereby aggregating global contextual information alongside localized details. We further introduce a feature homogenization module equipped with a set of learnable filters, improving the precision of multiscale feature fusion. Through extensive experiments and analysis on eight benchmark data sets, we show that MISU outperforms previous approaches by obtaining state-of-the-art performance, i.e., achieving a peak signal-to-noise ratio of 24.415 (versus 24.380) and an structural similarity index measure of 0.869 (versus 0.868).