Globally Deformable Information Selection Transformer for Underwater Image Enhancement

In the rapidly evolving image processing domain, transformers have emerged as powerful tools, yet significant challenges are encountered when they are applied to underwater image enhancement, such as visual disparity and computational inefficiency. Transformers do not have a unique module to maintain their performance while reducing the number of parameters. This study addresses the gap in the literature by introducing the globally deformable selection transformer (GS-Transformer), which is a model designed to enhance global feature selection and pixel connectivity, thereby reducing the computational complexity of the model while maintaining the image processing effect. Our novel multiresolution encoder-decoder module explicitly incorporates global information, overcoming the limitations of traditional transformers, whereas the multilocal coherence preserving loss (MCPL) mechanism ensures content integrity and coherence. Compared with the latest transform-based underwater image algorithms, this method is 15 times faster and utilizes only 41.7% (or approximately a half less) of the number of parameters. The experimental results on the UIEB, EUVP, and Synthesize datasets reveal that GS-Transformer achieves state-of-the-art performance in underwater image enhancement, with a reduced parameter number and improved efficiency, representing a significant advancement in the field. Our research will promote the application of the transformer in scenarios with high real-time performance.