Super-resolution reconstruction of turbine blade CT images based on realistic degradation modeling and enhanced Swin Transformer GAN

Abstract: To address the common issues of insufficient resolution, severe noise interference, and blurred fine structures in turbine blade cone-beam CT images, this study proposes an image super-resolution (SR) reconstruction method that integrates realistic degradation modeling with an enhanced Swin Transformer GAN. First, a realistic degradation model incorporating blur-kernel estimation and real-noise extraction is proposed to generate high- and low-resolution image pairs that better match the degradation characteristics of real cone-beam CT imaging. Second, an enhanced Swin Transformer network is designed to jointly capture global dependencies and local high-frequency details, while improving the representation of edge textures and gradient information. On this basis, an ESTGAN model is constructed by employing the enhanced Swin Transformer as the generator and a spectrally normalized U-Net as the discriminator. Extensive experiments on both simulated and real CT datasets demonstrate that the proposed method achieves superior quantitative and visual performance, particularly in terms of fine-structure recovery, edge preservation, and artifact suppression.