IM-CTSDG: Ancient Mural Image Inpainting Guided by Texture Structure and Multi-Scale Contextual Feature Aggregation

To achieve the interaction between structural and texture information during image restoration, enhance the semantic realism of irregular repairs, and improve restoration performance in cases of large area loss, this paper proposes the Image Restoration Network IM-CTSDG, built upon the original CTSDG algorithm. This network replaces all partial convolutions in the network generator with Fast Fourier Convolution (FFC) to address the issue of weakened capability to understand deep information when there is significant image damage. Designed the Receptive Field Expansion Module and Bi-directional Gated Feature Fusion (RFEM-BiGFF) to address local semantic loss and enhance global consistency. And we designed the Multi-Scale Contextual Feature Aggregation (MCFA) module to enhance the network’s feature extraction capabilities, capture long-distance information, and generate richer details. We applied the proposed method to a self-built ancient mural dataset with different damage ratios and conducted numerous experiments. The results demonstrate that for damage in different regions, the IM-CTSDG algorithm excels in mainstream image restoration algorithms, achieving superior performance in objective evaluation metrics such as L1, PSNR, SSIM, and MSE. Additionally, it exhibits higher semantic correctness and effectively restores the structure and texture information of damaged images, further validating the superiority of the proposed algorithm.