改进U-Net的全局特征融合水下图像增强网络

Key Points

• To develop an improved U-Net model that enhances underwater images by reducing color distortion and detail loss.
• Designed multi-residual convolution modules in the encoder and decoder for hierarchical feature fusion.
• Introduced channel attention modules in the decoder to weight channel importance and reduce varying degradation.
• Implemented convolution-permutation self-attention in the decoder to integrate global information for image reconstruction.
• Achieved PSNR of 23.42, SSIM of 0.9005, and LPIPS of 0.1385 on the UIEB dataset.
• Achieved PSNR of 29.35, SSIM of 0.9382, and LPIPS of 0.0880 on the LSUI dataset.
• Demonstrated effective recovery of color deviations and reduction of detail blur.