Underwater dam image enhancement based on CNN-transformer fusion

Area of Science:

• Hydropower Engineering
• Computer Vision
• Image Processing

Background:

• Underwater dam safety inspections require high-precision image analysis.
• Underwater images suffer optical degradation (noise, color shift, blur), hindering defect detection.
• Existing methods fail to fully address underwater optical characteristics and deep learning limitations.

Purpose of the Study:

• To propose an innovative image denoising and super-resolution network for underwater dam images.
• To address limitations of traditional and deep learning methods in restoring degraded underwater structural images.
• To improve the accuracy of dam defect detection and safety assessments.

Main Methods:

• Developed an Enhanced Super-Resolution Transformer GAN (ESRTGAN) integrating CNN and Vision Transformer.
• Employed multi-scale feature fusion, adaptive channel attention, and progressive training.
• Focused on fusing local feature extraction (CNN) with global context modeling (Transformer).

Main Results:

• ESRTGAN demonstrated excellent performance on real dam underwater image datasets, measured by PSNR, SSIM, and LPIPS.
• The network effectively restored image quality, preserving critical structural details and color.
• Achieved high computational efficiency while meeting manual interpretation standards.

Conclusions:

• ESRTGAN provides a robust solution for enhancing degraded underwater dam images.
• The method offers reliable technical support for automated analysis in long-term dam health monitoring.
• Improved image quality leads to more accurate safety assessments and defect detection.