关键词: 水下图像增强, Swin Transformer, 生成对抗网络, 多头自注意力机制, 马尔科夫判别器

Abstract: Addressing the issues of low contrast, heavy noise, and color distortion in underwater images, a Generative
Adversarial Network (GAN) model, is proposed as the core framework for underwater image enhancement. First, the generative
network follows an encoder-bottleneck-decoder structure, where the input feature maps are divided into multiple non-overlapping local
windows at the bottleneck layer. Second, a dual-path window multi-head self-attention mechanism is introduced to enhance local
attention while simultaneously capturing global information and long-range dependencies. Finally, the decoder recombines the multiple
windows back into the original size feature maps. The discriminator network employs a Markov discriminator. Compared to the
URSCT-SESR model, the proposed model shows an improvement of 0.8376 dB in Peak Signal-to-Noise Ratio (PSNR) and 0.0036 in
Structural Similarity Index (SSIM) on the UFO-120 data set. On the EUVP-515 dataset, the model achieves more significant
improvement, with a 0.8439 dB boost in PSNR, a SSIM increase of 0.0051, a 0.1124 enhancement in the Underwater Image Quality
Measure (UIQM), and a slight increase of 0.001 in the Underwater Color Image Quality Evaluation (UCIQE). Experimental results
demonstrate that the proposed model excels in both subjective and objective evaluation metrics, achieving notable improvements in
correcting color deviations in underwater images.

Key words: underwater image enhancement, Swin Transformer, Generative Adversarial Network (GAN), multi-head attention mechanism, patch discriminator(PatchGAN)