高级检索

中国古画渐进式多级特征修复算法

Progressive Multilevel Feature Inpainting Algorithm for Chinese Ancient Paintings

  • 摘要: 针对中国古画图像修复后的结果图像内容不一致且存在明显人工痕迹, 修复区域的细节信息丢失和修复后的内容模糊的问题, 提出利用绘画图像的多级语义特征来渐进式地修复中国古画. 将中国古画的修改过程分为5个不同的阶段, 分别从宏观、中观、微观的语义层面来渐进式修复修复古画图像. 修复模型首先关注高抽象级别的语义信息和大尺度的结构信息, 然后逐渐地将注意力转移到越来越精细的尺度上, 而不必同时学习所有尺度的信息, 由古画的高层的语义特征渐进式地向较低层的语义特征的修复. 该方法不仅使得修复过程更加有序, 而且在次序靠后的每个阶段能够利用先前阶段中学到的知识来简化后续阶段的修复. 依据11位作家共197幅高清数字化古画制作了山水画、街景画、花鸟画、人物画等4个数据集, 在制作的数据集上进行的实验表明本文算法在PSNR, SSIM, IS, FID等指标上优于一步到位的修复算法.

     

    Abstract: There are some problems in the field of Chinese ancient painting images, such as inconsistent repaired image contents and obvious artificial traces, loss of detailed information in the repaired area, and blurred repaired content. To address these problems, we propose to use the multi-level semantic features of the painting image to repair ancient Chinese paintings gradually. The revision process of ancient Chinese painting is divided into five different stages, which are to repair the ancient painting image gradually from the macro, meso, and micro semantic levels. The repair model first pays attention to the high-level semantic information and large-scale structural information and then gradually shifts its attention to more and more fine scales without learning the information of all scales at the same time. The repair model gradually changes from the high-level semantic features of ancient paintings to the low-level semantic features. This method not only makes the repair process more orderly but also makes use of the knowledge learned in the previous stage to simplify the repair in the subsequent stage. Based on 197 high-definition digital ancient paintings by 11 writers, four data sets of landscape painting, street view painting, flower, and bird painting, and figure painting are made. Experiments on the above data sets show that the algorithm in this paper is better than the one-step repair algorithms in PSNR, SSIM, IS, FID.

     

/

返回文章
返回