Department of Computer Science and Technology, Xinzhou Normal University, Xinzhou, China.
School of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan, China.
PLoS One. 2024 Sep 6;19(9):e0307811. doi: 10.1371/journal.pone.0307811. eCollection 2024.
The current mainstream image restoration methods have difficulty fully learning the structure and color information of murals in mural image restoration tasks due to the limited size of the available datasets, resulting in problems such as structural loss and texture errors. This study proposes a two-stage mural restoration network based on an edge-constrained attention mechanism. This paper introduces additional sketches as inputs during the coarse restoration phase and incorporates a local edge loss function to enable the network to generate corresponding structural information based on the sketches. In the fine restoration phase, the calculation for the similarity between missing areas and known areas is optimized to enhance the consistency of the restoration results with the texture of the known areas. Furthermore, a structure-guided attention propagation block is introduced after adopting the attention mechanism. This block selectively integrates surrounding contextual information to update the attention score map, thereby enhancing the coherence and plausibility of the generated textures. The experimental results show that the proposed method outperforms the current mainstream restoration methods according to various assessment indices. The proposed method generates high-quality structural information according to user guidance information, and the repaired texture is highly visually consistent with that of the original mural, with few noticeable deviations. This study provides a new approach for mural restoration, which may positively impact cultural heritage protection and artistic restoration applications.
当前主流的图像恢复方法由于可用数据集的规模有限，在壁画图像恢复任务中难以充分学习壁画的结构和颜色信息，导致结构丢失和纹理错误等问题。本研究提出了一种基于边缘约束注意力机制的两阶段壁画恢复网络。本文在粗恢复阶段引入额外的草图作为输入，并结合局部边缘损失函数，使网络能够根据草图生成相应的结构信息。在精细恢复阶段，优化了缺失区域和已知区域之间的相似性计算，以增强恢复结果与已知区域纹理的一致性。此外，在采用注意力机制后引入了结构引导的注意力传播块。该块选择性地整合周围的上下文信息来更新注意力得分图，从而增强生成纹理的连贯性和真实性。实验结果表明，根据各种评估指标，所提出的方法优于当前主流的恢复方法。该方法根据用户指导信息生成高质量的结构信息，修复后的纹理与原始壁画高度视觉一致，几乎没有明显的偏差。本研究为壁画修复提供了一种新方法，可能对文化遗产保护和艺术修复应用产生积极影响。
