To address the limitations of traditional restoration techniques regarding structural continuity and color consistency, this paper introduces a two-stage generative adversarial network (GAN)-based method for mural restoration. The approach comprises two phases: the first phase involves the adversarial training of an edge generator G1 and a discriminator D1 to accurately generate the edge structure of obscured areas; the second phase utilizes a structure reconstruction network G2 with an encoder-decoder architecture, enhanced with multi-level skip connections and instance normalization techniques, to restore the detailed textures of damaged regions. Experiments were conducted on the DhMurals1714 dataset and a self-constructed dataset of Han Dynasty murals. The outcomes show that our approach surpasses current methods in terms of PSNR and SSIM metrics when the mask rate is between 10% and 40%. Ablation studies have confirmed the effectiveness of the edge generation network, skip connections, and the improved loss function, particularly under high mask rates (30%–40%), where the complete model achieved a 1.59 dB PSNR improvement over the baseline model. This method demonstrates significant advantages in maintaining the structural integrity and natural color transitions of murals, offering an effective solution for the digital preservation of cultural heritage.