基于字形增强和Transformer的损坏手写东巴字符图像修复

Text Image Inpainting for Damaged Handwritten Dongba Characters via Glyph Enhancement and Transformer

摘要: 受自然老化、环境侵蚀等因素影响，古籍普遍存在文字损坏的现象，现有通用图像修复算法因使用非专用的掩码且难以理解文字结构，导致出现算法泛化性不足和字形结构修复错误问题。针对上述问题，构建了一个模拟真实损坏的专用掩码数据集(其中包含九种损坏类型)，同时提出了一种融合字形增强和Transformer的东巴字符图像修复算法。该算法首先引入字形增强模块，利用小波分解在低频域提取字形结构特征信息；然后，利用字形结构特征信息引导Transformer对缺失区域进行结构推断，重建结构完整、纹理清晰的东巴字符图像。在损坏手写东巴字符图像数据集上的对比实验表明，所提方法在不同掩码比例上在多项指标上均优于现有算法。实验结果验证了所提算法的有效性，并为古籍文字及少数民族象形文字的修复提供了可行方案。

Abstract: Influenced by factors such as natural aging and environmental erosion, ancient documents universally suffer from text degradation. Existing general inpainting algorithms, due to the use of non-specialized masks and a failure to capture character structures, result in weak generalization and structural errors in reconstructed glyphs. To address these issues, we construct a specialized mask dataset simulating real-world damage (encompassing nine damage types) and propose a Dongba character inpainting algorithm that integrates a glyph enhancement module with a Transformer. The algorithm first employs the glyph enhancement module, which utilizes wavelet decomposition to extract structural feature information from the low-frequency domain. Subsequently, these structural features guide the Transformer to perform structural inference and content generation for missing regions, thereby reconstructing Dongba character images with complete structures and clear textures. Comparative experiments on a damaged handwritten Dongba character dataset demonstrate that the proposed method significantly outperforms existing algorithms across multiple metrics at various mask ratios. Experimental results validate the effectiveness of the proposed algorithm and offer a viable solution for the inpainting of ancient texts and ethnic minority pictographic scripts.