High-resolution images of the retinal fundus are important for the diagnosis of diabetic retinopathy, but clinical data often contain low-quality images that obscure the fine vascular structures necessary for accurate diagnosis. Existing super-resolution methods have fundamental shortcomings: convolutional neural networks produce overly smooth results, while generative adversarial networks can produce spurious artifacts. The research team proposed a new method called Dual-Transformer Residual Super-Resolution Network (DTRSRN), which combines Swin Transformers to process the overall context with a parallel network to preserve the details of vascular structures. A key innovation is the use of fractal dimension analysis to measure how well the morphological properties of vessels are preserved. The results showed that DTRSRN achieves 33.64 dB PSNR at twice the resolution increase, outperforming other methods including SwinIR, HAT, and ResShift. The method achieved a 17.0% improvement in the preservation of vascular structure compared to the best comparative method, which has the potential to improve the diagnosis of eye diseases.