Addressing the issues of artifacts and insufficient contextual awareness in existing algorithms for repairing severely damaged facial images, we propose a facial image restoration method based on global-local feature fusion. First, a wavelet-masked shuffling down-sampling module is employed to enhance the model's capability to learn local features of edge textures, thereby resolving the issue of inadequate extraction of facial local details during the restoration process. Secondly, a global channel-weighted attention mechanism is designed to extract global features, enabling the model to focus specifically on the most critical feature channels for the current task, effectively reducing unnecessary computational overhead. To ensure the importance of achieving high-quality output, the model selectively filters and adjusts information flow. Lastly, a multi-scale pooling module is utilized to adaptively fuse the extracted features, allowing the model to better filter out noise and preserve useful signals, thereby enhancing the algorithm's applicability and robustness in complex environments.Through end-to-end learning, the model optimizes both global and local features simultaneously, enriching the final feature maps with richer semantic information. Experimental results validated on the CelebA-HQ high-resolution facial dataset demonstrate our method's superiority in terms of clarity and coherence over comparative methods in qualitative experiments. Quantitative experiments show significant advantages in Structural Similarity Index (SSI), Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS), and L1 loss metrics.In summary, the proposed method shows superior performance in the task of facial image restoration.