Defect Repair of Murals Guided by Fusion Structural and Textural Feature

ABSTRACT To address the challenge of insufficient restoration accuracy for large‐area damaged murals, this paper proposes a two‐stage generative adversarial network model based on Fourier convolution, employing a progressive coarse‐to‐fine restoration strategy. In the coarse restoration stage, the Fourier convolution module endows the network with high‐dimensional perceptual capability through frequency‐domain feature extraction, enabling rapid inpainting of damaged regions. Simultaneously, an edge detection module accurately captures the structural contours of the coarse restoration results, providing critical cues for structure decoding in the subsequent fine restoration stage. In the fine restoration stage, a dual‐generative adversarial network model is guided by the synergistic interaction of structure and texture features, followed by a feature fusion module to achieve feature reconstruction. Through cascaded optimization of the two stages, the restoration results show significant improvements in structural coherence and texture authenticity. Experimental results demonstrate that, compared with several existing mainstream methods, the proposed model achieves outstanding performance in terms of peak signal‐to‐noise ratio (PSNR) and structural similarity index, reaching up to 33.92 dB and 0.95, respectively, fully validating its effectiveness in complex large‐area damage restoration scenarios. To address the problem of insufficient repair accuracy for large‐area damage of wall paintings, a two‐stage generative adversarial network model based on Fourier convolution is proposed, which adopts the progressive strategy of two‐stage coarse‐fine repair.