High-resolution ultrasonic image reconstruction of shallow rail defects using a Relativistic-Attention-Weighted CycleGAN

To address the challenges of low imaging resolution and missed detection of rail Shallow Surface Layer (SSL) defects – such as holes and cracks – in ultrasonic phased array testing, this paper proposes a high-resolution image reconstruction method based on Relativistic-Attention-Weighted CycleGAN (RAW-CycleGAN). An image weighting module embedded in the input channel enhances weak pixel features, while a spatial attention mechanism integrated into the generator strengthens the extraction of spatial structural features from ultrasonic images. Additionally, the traditional PatchGAN discriminator is replaced with a relativistic discriminator to improve network training stability. Experiments were conducted on typical defects, and the quantitative analysis of detection results was performed from three perspectives: image reconstruction similarity (MSE, PSNR, SSIM, Perceptual Loss), shape reconstruction accuracy (circularity or rectangularity), and weak feature extraction capability (missing detection rate). Compared with the total focusing method, RAW-CycleGAN achieves a 93.72% reduction in MSE, a 92.49% increase in PSNR, a 48.53% improvement in SSIM, a 93.67% decrease in Perceptual Loss, a 96.43% enhancement in circularity, and a 25% reduction in the missing detection rate. These evaluation metrics indicate that compared to other enhanced CycleGAN models, RAW-CycleGAN can significantly improve the reconstruction accuracy of SSL defects and maintain high training stability.