摘要： 针对敦煌壁画修复过程中上下文特征提取不充分、图像特征信息丢失的问题，提出了一种基于高效特征表达的敦煌壁画修复方法。该方法充分结合注意力机制与多尺度特征融合，以提升破损壁画的修复质量。首先，在编码-解码结构的基础上，引入双注意力模块在壁画图像的空间和通道维度上进行特征细化，以增强壁画的上下文特征表达能力，提高修复区域的语义一致性与细节还原效果；然后，在跳跃连接处引入多尺度门控模块，以捕获编码阶段的特征信息并传递至解码器，从而增强特征的长程依赖性，提升全局结构与局部细节的融合能力。最后，为降低计算复杂度并最大限度地保留有效特征信息，提出了非线性激活自由块，实现优化模型的计算效率和修复质量。为了验证所提方法的有效性，在敦煌壁画数据集和FFHQ人脸数据集上进行了实验。实验表明，所提方法不仅适用于壁画修复任务，在其他类别的图像修复任务中同样具有良好的泛化能力，能够生成更自然的修复结果，评价指标优于所提其他算法。

Abstract: To address the challenges of insufficient contextual feature extraction and information loss in Dunhuang mural inpainting, this paper proposes an inpainting method based on efficient feature representation. The proposed method integrates attention mechanisms with multi-scale feature fusion to enhance the quality of inpainted murals. Specifically, built upon an encoder-decoder architecture, a dual attention module is introduced to refine features in both spatial and channel dimensions, thereby enhancing the contextual feature representation of murals and improving semantic consistency and detail reconstruction in damaged regions. Furthermore, a multi-scale gating module is incorporated into the skip connections to capture and transmit feature information from the encoding stage to the decoding stage, thereby strengthening long-range dependencies and improving the fusion of global structures with local details. To further reduce computational complexity while preserving effective feature information, a nonlinear activation-free block is proposed, optimizing both computational efficiency and inpainting quality. To evaluate the effectiveness of the proposed method, extensive experiments are conducted on the Dunhuang mural dataset and the FFHQ face dataset. Experimental results demonstrate that the proposed method not only performs well in mural inpainting tasks but also exhibits strong generalization ability across different image inpainting tasks, generating visually natural inpainting results and achieving superior performance in comparison to existing algorithms.