基于U-Net的历史文档图像分割研究  

Image Segmentation of Historical Documents Based on U-Net

作　　者：贾晓光 伍星[1] 何中市[1] 但静培[1] JIA Xiao-guang;WU Xing;HE Zhong-shi;DAN Jing-pei(School of Computer,Chongqing University,Chongqing 400000)

机构地区：[1]重庆大学计算机学院,重庆400000

出　　处：《现代计算机》2020年第19期52-56,共5页Modern Computer

基　　金：国家自然科学基金资助项目(No.51608070)。

摘　　要：文档图像分割是历史文档分析的关键技术支撑。针对历史文档布局复杂、页面黄化等造成其分割难的特点,提出基于U-Net的端到端的像素级别历史文档图像分割模型。该模型以嵌入空洞卷积的DenseNet为骨干网融合更丰富语义特征来解决历史文档图像布局复杂问题,使用Focal Loss函数解决数据集中类别极不均衡问题。在数据集DIVAHisDB的实验表明,该方法相对之前历史文档图像分割方法有更好的效果。Image segmentation of historical documents is the key technical support for historical document analysis.Aiming at the characteristics of complex historical document layout and yellowed pages,the end-to-end pixel-level historical document image segmentation model based on U-Net is proposed.In our method,DenseNet with dilated convolution is applied as the backbone network to provide richer semantic features,solving the problem of complex layout in historical documents.We introduce Focal Loss function to solve the problem of extremely imbalanced categories for the dataset.Experiments on the DIVA-HisDB show that the method we proposed has better results than previous methods for page segmentation of historical document images.

关 键 词：历史文档 U-Net DenseNet Focal Loss 空洞卷积 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]