RetUIE: Retention-based Underwater Image Enhancement

This is the authors' abstract. We don't add key points for this paper.

Recently, Transformer have been widely applied and have demonstrated superior performance on computer vision important tasks such as underwater image enhancement and image dehazing. However, Transformers also face challenges of excessive memory usage and limited parallel training that prevent image restoration on resource-constrained devices. To address the problem, retention mechanism-based method (RetNet), which is a parallelizable model with recurrent formulations that enables parallelization of training, low-cost inference and good performance, is proposed in natural language processing. Inspired by RetNet, the retention mechanism-based network for underwater image enhancement is proposed (RetUIE). Specifically, considering color channel loss in underwater images, a retention convolution channel self-attention module is introduced. Further, the retention shifted window pixel self-attention module has been proposed to address the problems of non-uniform image degradation and improve computational efficiency. Finally, underwater images are restored through residual connections. Experimental results on multiple public datasets demonstrate that the proposed RetUIE outperforms state-of-the-art methods both qualitatively and quantitatively. Notably, on the UIEB dataset, RetUIE achieves a performance milestone with a PSNR of 26.3 dB and SSIM of 0.93. Moreover, RetUIE can be extended to image dehazing. The code will be released on GitHub after acceptance.