EdgeConnect model based on Fourier convolution for Dunhuang mural costume image restoration

Image inpainting is crucial for image editing and cultural heritage restoration. The complex structures and large damaged areas of Dunhuang murals make traditional methods ineffective for recovering detailed textures and natural appearance. This study proposes an improved mural costume restoration method based on the EdgeConnect model enhanced with Fast Fourier Convolution (FFC). By replacing standard downsampling convolutions in the first-stage generator with FFC modules, the model incorporates frequency-domain information to improve edge and texture consistency. Spectral Transform operations capture both global and local features, enabling smoother transitions and finer detail restoration. The global receptive field of FFC also enhances robustness to irregular masks and large occlusions. Experiments show that the proposed method significantly outperforms existing techniques in restoring highly damaged murals. Quantitative metrics such as SSIM and PSNR confirm its superiority. This approach provides theoretical and technical support for high-quality image restoration in cultural heritage applications.