To address the significant decline in face recognition performance caused by low resolution, high noise, and complex degradation factors in security surveillance scenarios, this paper proposes a joint optimization framework that integrates a Transformer and a Generative Adversarial Network (GAN). The innovation of this framework lies in: (1) designing the Face Reconstruction Transformer (FRFormer), which integrates a hierarchical window attention mechanism and a multi-level feature pyramid structure, enhancing the ability to retain identity features through local-global collaborative modeling; (2) constructing the GFP-GAN reconstruction model, which combines pre-trained face priors and degradation removal modules, and utilizes adversarial training to improve image authenticity and detail restoration. Experiments show that when the input is 32 × 32 pixels, the PSNR of GFP-GAN is increased by more than 8 dB, and the SSIM reaches 0.953; FRFormer achieves recognition accuracies of 99.58% and 96.31% on the LFW and AgeDB-30 benchmarks, respectively, which are 0.08 and 0.13 percentage points higher than those of Swin Transformer. Ablation experiments verify the effectiveness of the window attention mechanism and hierarchical reconstruction strategy, especially in noise suppression and cross-pose recognition tasks. This framework has broad application potential in degraded visual conditions, such as biometric recognition and medical image analysis, and provides an end-to-end solution for low-quality face recognition.