边缘引导和拉普拉斯金字塔分解的古文本图像修复算法
An Ancient Text Image Inpainting Algorithm via Edge Guide and Laplacian Pyramid Decomposition
刘畅 1张玲 1何英豪1
作者信息
- 1. 武汉科技大学计算机科学与技术学院 武汉 430065;武汉科技大学大数据科学与工程研究院 武汉 430065;智能信息处理与实时工业系统湖北省重点实验室 武汉 430065
- 折叠
摘要
针对当前图像修复算法应用到古文本图像上时,出现纹理模糊或结构内容不完整的问题,提出边缘引导和拉普拉斯金字塔分解的古文本图像修复算法.首先利用边缘修复模块对古文本图像的边缘结构进行修复,重建缺损区域的边缘信息;然后利用预训练的文字学习模块对局部缺损区域进行内容修复,得到一幅局部内容修复图像,并进行拉普拉斯分解;最后在拉普拉斯金字塔修复模块中,根据图像的低层和高层特征,利用内容修复模块对图像进行递进修复,内容修复模块中引入双交叉编码器和多尺度融合块,有助于获取更加有效的特征信息,生成纹理结构完整的图像修复结果.在古文本图像数据集的测试集上进行实验的结果表明,各项图像质量评估指标中,峰值信噪比为34.322dB,结构相似性为0.970,均方根误差为5.203,验证了所提算法的有效性和可行性.
Abstract
Current image inpainting methods are often perform poorly on ancient text images,producing re-sults with blurred textures or incomplete structural content.To address this problem,we propose an inpaint-ing algorithm for ancient text images via edge guide and laplacian pyramid decomposition.We first use an edge restoration module to restore the edge structure for the damaged regions and construct an edge-guided map.Then,we employ the pre-trained text learning module to restore the local damaged regions and obtain a local inpainting image,which is decomposed into a content image and a detail map through laplacian pyra-mid transform.At last,in the Laplace pyramid restoration module,the content restoration module is used to progressively repair the image according to the low-level and high-level features of the image.The content restoration module introduces a dual cross-encoder and multi-scale fusion blocks to prompt the module to obtain more effective feature information and generate desirable image inpainting results.The superiority quantitative results on the benchmark dataset demonstrate the effectiveness and feasibility of the proposed method,that peak signal to noise ratio(PSNR)is 34.322 dB,structural similarity(SS1M)is 0.970 and root mean square error(RMSE)is 5.203.
关键词
图像修复/古文本图像/边缘图/双交叉编码器/多尺度融合块Key words
image inpainting/ancient text images/edge images/dual cross-coders/multiscale fusion blocks引用本文复制引用
出版年
2024
NETL
NSTL
万方数据