Abstract
In the field of hydropower engineering, the safety inspection of underwater dam structures relies heavily on high-precision image analysis. However, images captured by underwater robots generally suffer from optical degradation issues, such as speckle noise interference, blue-green color shift, low contrast, and blurred details. Traditional image denoising algorithms fail to fully account for the optical transmission characteristics of the underwater environment, often leading to the loss of critical structural detail information when processing dam underwater images. Meanwhile, existing deep learning-based image processing methods, which lack comprehensive modeling of the physical process of underwater image degradation, struggle to achieve ideal results in color restoration and edge preservation. These image quality issues may cause misjudgments in dam defect detection, seriously affecting the accuracy of safety assessments. To address the above problems, this study proposes an innovative image denoising and super-resolution network, namely Enhanced Super-Resolution Transformer GAN (ESRTGAN), which integrates CNN-based local feature extraction and Transformer-based global context modeling to tackle dam underwater image degradation. By effectively fusing the local feature extraction capability of Convolutional Neural Networks (CNN) and the global context modeling capability of Vision Transformer, ESRTGAN achieves high-quality image restoration and enhancement. The network adopts multi-scale feature fusion strategy, adaptive channel attention mechanism, and progressive training method, significantly improving the reconstruction capability of image details. Experimental results show that ESRTGAN exhibits excellent performance in terms of PSNR, SSIM, and LPIPS metrics in the enhancement experiments on real dam underwater image datasets while maintaining high computational efficiency. Additionally, the processed images meet the requirements of manual interpretation in subjective visual effects, providing reliable technical support for the automated analysis of long-term dam health monitoring.