基于Huber函数双边全变分的多帧文档图像超分辨率重建

针对低分辨率文档图像中噪声模型不确定、字符边缘和纹理走向复杂多变的问题,提出GemanMcClure(GM)范数替代L1、L2范数用于提高算法的鲁棒性,设计了结合双边全变分(BTV)和Huber函数的正则化项,采用Lucas-Kanade光流配准算法,利用字符结构特征的先验信息,使算法在重建过程中更加注重边缘细节与边缘方向信息。实验表明,与L1BTV、L2BTV和无Huber函数的GMBTV正则化(下文简称GM方法)重建方法相比,文中算法在混合噪声模型下能够显著平滑噪声、锐化边缘、提升文档图像字符的分辨率,字符识别率提高14.69%的同时运算时间缩短了29.34%。

The Super-resolution Reconstruction of Multi-frames Document Images Based on Huber Function and Bilateral Total Variation

It has been a key problem in document image super resolution which the noise model is uncertain, edge and texture of characters tend to complex and changeable. Geman&McClure (G&M) norm instead of L1 and L2 norm was proposed and used to improve the robustness of the algorithm. A regularization item combining Bilateral Total Variation (BTV) and Huber function was designed with adopting Lucas-Kanade light flow registration algorithm which is highly compatible for the proposed algorithm. This algorithm uses full information of the characters structure characteristics to make the algorithm pays more attention to the edge details in the process of reconstruction, applies the edge direction information more efficiently, and promotes information fusion of a series of low resolution images. Finally, experiments corresponding to L1BTV, L2BTV and G&M BTV regularization algorithm without Huber function were carried out, the results show that the algorithm under the G&M norm with combination of BTV and Huber function is much more excellent than other algorithms. In the environment of mixed noise model, document image super-resolution reconstruction can smooth the noise significantly, sharpen the edge and improve the resolution of characters in document image. The recognition rate of characters was improved by 1469%, and the operation time of the proposed algorithm was shorten by 29.34% at the same time.