基于MD-LinkNet的低质量文档图像二值化算法

Binarization of degraded document images based on MD-LinkNet neural networks

针对低质量文档图像存在的背景渗透、页面污渍、边缘大面积与文本相似的噪声等现象,改进D-LinkNet框架,提出了一种融合多尺度特征(multiple scale feature)的低质量文档图像二值化算法,简称为MD-LinkNet。该算法有两处改进,一是在编解码中间部分增加剩余多核池化(RMP)模块来通过四个池化操作以提取丰富的文档特征信息;二是将池化后的低分辨率图像通过DUpsample而不是双线性插值进行上采样,结合了文档图像像素邻域信息,将文档图像的全局与局部特征进行融合,提高了分割精度。实验结果表明,在2017年和2018年国际文档图像二值化竞赛(DIBCO)数据集中,本文算法的F值(F-measure)最高分别达到了90.54、91.42,验证了所提出算法在解决多种复杂噪声背景的低质量文档图像下的鲁棒性,且相比其他最新经典算法效果较优。

Aiming at the bleed through,page stains,text-like noise on the large area edges and so on,which occurred in the low quality document images,the D-LinkNet framework is improved,and a low quality document image binarizatin algorithm combining with multiple scale features is proposed.We referred to as MD-LinkNet.The algorithm has two improvements based on the original network.One is to add the remaining multi-core pooling(RMP)module in the middle of the codec to extract rich document feature information through four pooling operations;the second is to pass the pooled low-resolution image through DUpsample instead of bilinear interpolation performs upsampling,not only combines the pixel image neighborhood information of the document image,but also to fuses the global and local features of the document image to improve the segmentation precision.The experimental results show that in the 2017 and 2018 International Document Image Binarization Competition(DIBCO)dataset,the F-measure of the algorithm reaches 90.54 and 91.42 respectively,the results of extensive experiments on many datasets show the robustness of the proposed technique on various types on degradations in the document images where our technique demonstrates superior performance against many other state of art classcal methods.