边缘信息与线描稿引导下的敦煌壁画多尺度特征修复网络OA

To solve the problems of unreasonable repair structure and inconsistent texture inpainting that arise in the pro-cess of restoring Dunhuang mural images,this paper proposes a multi-scale feature extraction inpainting network that combines edge information and line drawings for guidance.In order to facilitate a rational repair of the mural structure,the paper designs a first stage edge information prediction network,in which the edge information extracted by the Canny algorithm and the line drawing information are input into a predictive network that combines linear deformable convolu-tion with a residual network for edge extraction predictions,resulting in an edge prediction map that can guide the second stage of image inpainting.To enhance the model's understanding of texture details in the second stage image inpainting network,a residual network is used to preserve mural features,while a multi-scale feature extraction module with gating units is designed to extract and filter mural information,which is then fused to complete the image repair.By comparing the results with four other algorithms in the mold spot mask repair experiment,the PSNR value improves by about 5.457～13.411 dB,the SSIM increases by about 0.041～0.077,and the L2 loss decreases by about 0.004～0.02.The experimental results indicate that the proposed model outperforms the comparison experiments in both subjective and objective evalua-tion metrics,achieving more rational inpainting results in terms of structure and texture.