An efficient swin transformer-based method for underwater image enhancement

Due to the complex imaging environment of the ocean, the underwater images obtained by optical vision systems are usually severely degraded. Recently, methods for enhancing underwater images are mostly based on deep learning. However, the intrinsic locality of convolution operation makes it difficult to model long-range dependency efficiently, which may lead to the limited performance of these methods. This paper proposes an efficient method for underwater image enhancement by utilizing Swin Transformer for local feature learning and long-range dependency modeling. The network structure of this method is mainly composed of encoder, decoder and skip connections, in which the encoder and decoder take the Swin Transformer block as the basic unit. Specifically, the encoder is used to learn multi-scale feature representations, and the decoder is utilized to upsample the extracted contextual features progressively. Skip connections are used to fuse multi-scale features from the encoder and decoder. Experimental results demonstrate that the proposed method outperforms state-of-the-art methods on different datasets by up to 1.09 similar to 1.64dB (PSNR) and 1.9%similar to 2.3% (SSIM) in objective metrics, and achieves the best visual effect in subjective comparisons, especially in terms of color cast removal and sharpness enhancement.