基于生成对抗网络的文物图像修复与评价

目的 为了有效解决文物图像的不易保存和物理方法修复困难等问题，提出一种基于生成对抗网络的图像修复算法。方法 文中算法主要分为2个阶段，第1阶段通过Canny边缘检测器提取图像已知部分的边缘信息，利用1个生成器和1个鉴别器修复图像缺失边缘。第2阶段将第1阶段生成的边缘作为先验信息，通过1个生成器和2个鉴别器修复图像缺失部分。2个鉴别器由整体鉴别器和局部鉴别器组成，整体鉴别器用来评估修复后的图像整体连贯性，局部鉴别器用来查看待修复区域为中心的小区域局部一致性。结果 与传统算法对比，文中算法在提高生成图片纹理质量的基础上保证了全局语义结构一致性，在客观指标（峰值信噪比和结构相似性）上，较其他方法有更好的效果。结论 文中算法可以有效修复文物图像的缺损部分，尤其是结构复杂的大范围缺失，取得了良好的视觉效果，表明该算法有良好的修复性能。

Repair and Evaluation of Cultural Relic Images Based on Generative Adversarial Network

The work aims to propose an image repair algorithm based on generative adversarial network, so as to ef-fectively solve the difficulties in preserving cultural relic images and repairing them in physical method. The proposed algorithm was mainly divided into two stages. In the first stage, the edge information of the known part of the image was extracted by Canny edge detector, and a generator and a discriminator were used to repair the missing edges of the image. In the second stage, the edges generated in the first stage were taken as prior information, and the missing part of the image was repaired by one generator and two discriminators. The two discriminators consisted of a global discriminator and a local discriminator. The global discriminator was used to evaluate the overall coherence of the repaired image. The local discriminator was used to check the local consistency of a small area centered on the area to be repaired. Compared with the traditional algorithm, the proposed algorithm ensured the consistency of the global semantic structure by improving the quality of the generated image texture. In terms of objective indicators (peak signal-to-noise ratio and structural similarity), the results of the proposed algorithm were better than those of other methods. The algorithm proposed herein can effectively repair the defective parts of cultural relic images, especially large-scale missing complex structures, and has achieved good visual effects, indicating that the algorithm has good repair performance.