Super-Resolution Study of Infrared Images of Embankment Dam Leakage Based on Deep Learning
Fan Zhu, Yanchang Gu, S Wang, Wenzhong Zhou · 2024
Aiming at the problem of difficult recognition of leakage due to large detection distance and low imaging resolution in the infrared nondestructive detection of embankment dam leakage, an improved Enhanced Super-resolution Generative Adversarial Network (ESRGAN) tailored for super-resolution of leakage infrared images is proposed. The Efficient Channel Attention mechanism and feature fusion mechanism were integrated into the ESRGAN to improve super-resolution image clarity. Leakage infrared image super-resolution experiment based on the datasets created by the leakage infrared detection experiment of an earth dam model. Training result indicated that the improved ESRGAN achieved a peak signal-to-noise ratio (PSNR) of 29.6893 dB, representing an improvement of 18.60% compared to Bicubic, 14.47% over the Super-Resolution Convolutional Neural Network (SRCNN), 12.34% above the Super-Resolution Generative Adversarial Network (SRGAN), and 3.18% more than ESRGAN. Similarly, its structural similarity index (SSIM) reached 0.8897, representing a 22.40% increase over Bicubic, 19.61% over SRCNN, 8.26% over SRGAN, and 0.78% over ESRGAN. The improved ESRGAN produces images with greater clarity, providing valuable experience and reference for enhancing the accuracy and efficiency of infrared detection of embankment dam leakage.