Super-resolution Reconstruction of DR Images of Aerospace Turbine Blades Based on Swin Transformer

Microcracks near air-film pores on aviation turbine blades are a primary cause of blade failure. Current digital radiographic (DR) imaging techniques for detecting such microcracks encounter limitations, including artifacts, jagged edge in reconstructed images, and insufficient resolution to reliably resolve cracks smaller than 1 mm. A super-resolution reconstruction algorithm for digital radiographic images based on window transformation and self-attention network (SwinDR) has been established. The algorithm comprises three modules: shallow feature extraction, deep feature extraction, and upsampling reconstruction. During network training, perceptual loss and L1 loss constraints are employed to optimize image generation. Compared with existing methods that including the Enhanced Deep Super-resolution Network (EDSR), Residual Channel Attention Network (RCAN), and Swin Transformer-based image restoration model (SwinIR), the SwinDR algorithm exhibits superior performance. At 2x magnification, SwinDR-reconstructed DR images achieved an average peak signal-to-noise ratio (PSNR) of 39.724, and a structural similarity index measure (SSIM) of 0.947. At 4x magnification, the PSNR and SSIM values were 39.231 and 0.933, respectively. Reconstructed cracks measured as small as 0.452 mm in length and 0.031 mm in width. Results indicate that SwinDR effectively minimizes deformation and edge artifacts, enhances image contrast, restores the high-frequency details, and significantly improves reconstructed image clarity compared to conventional approaches.