基于文本特征自适应正则化的文档图像超分辨率重建

摘要

In bilateral total variation (BTV) regularized super-resolution reconstruction,the BTV regularization term can not adaptively perform smoothing to the characters in document images,which can not fully retain the image details.The modified BTV regularization term based on the local width and local direction of the character strokes was designed with fully use of the prior information of the characters in document images.By analyzing the input LR and its intermediate interpolation image,the local stroke width and stroke direction in the HR image were approximated.This information was encapsulated into the regularization term.By minimizing the linear combination of the regularization and data-fidelity terms,the HR image was reconstructed.The results compared with BTV regularization reconstruction methods show that the algorithm under the new regularization item is much more excellent.The document image super-resolution reconstruction can smooth the noise significantly while keeping the edge details.Reconstruction text strokes lines are smoother and stroke contour is clear.The resolution of characters in document images is enhanced in which the average PSNR is improved by 8.32％,and the average MSSIM is increased by 5.94％,the running time is reduced 26.2％.%在BTV正则化超分辨率重建中,由于BTV正别项无法根据字符特点自适应进行平滑滤波,不能充分保留图像细节的缺点,设计了基于文字笔画局部宽度和方向自适应的改进BTV正则化项提高字符分辨率,通过分析输入的LR图像及其中间插值图像,得到近似于HR图像中的局部笔画宽度和方向信息,并将此信息封装进正则化项,通过最小化正则化项和数据保真项的线性组合,重建高分辨率图像.实验结果表明,与原BTV正则化重建方法相比,算法在保留细节信息,提高字符分辨率的同时能够显著滤除噪声,重建得到的文字笔画线条较为平滑,且轮廓清晰,提升了文档图像的分辨率.其中平均PSNR提高了8.32％,平均MSSIM提高了5.94％,同时运行时间减少了26.2％.