Laser Ultrasonic Image Enhancement and Defect Automatic Detection Method for Cable Defects Based on Generative Adversarial Networks

Although laser ultrasonic imaging technology has the advantages of high resolution and non-contact detection, in practical application, the image is often affected by noise and equipment limitations, which makes it difficult to identify small defects. For this reason, this study combines the generative adversarial networks (GAN) and Convolutional Neural Network (CNN) to construct an end-to-end image enhancement and defect detection framework. In the image enhancement stage, the U-Net structure is used as the generator to optimize the image quality by fighting against training and perceptual loss, and the total variation (TV) regularization term is introduced to suppress noise, which significantly improves the image clarity and detail retention ability. In the part of defect detection, the feature extractor based on ResNet-18 architecture combined with transfer learning strategy effectively solves the problem of small sample training, and realizes the accurate location of defects through Grad-CAM thermography technology. The experimental results show that this method is superior to traditional methods and existing deep learning models in image enhancement index, and also shows significant advantages in defect detection performance, especially in the identification and location of tiny defects.