联合度量指标损失和U-Net的文档图像二值化  

Document image binarization via combining measure index loss and U-Net

作　　者：张晶 周稻祥 吴永飞 冯姝 ZHANG Jing;ZHOU Dao-xiang;WU Yong-fei;FENG Shu(College of Mathematics,Taiyuan University of Technology,Jinzhong 030600,China;College of Data Science,Taiyuan University of Technology,Jinzhong 030600,China;Department of Foundation,Shanxi Agricultural University,Taigu 030801,China)

机构地区：[1]太原理工大学数学学院,山西晋中030600 [2]太原理工大学大数据学院,山西晋中030600 [3]山西农业大学基础部,山西太谷030801

出　　处：《计算机工程与设计》2024年第8期2400-2406,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(62201331、62101376、61901292);山西省应用基础研究计划基金项目(201901D211078、20210302124543)。

摘　　要：当前深度神经网络模型在图像分割时均采用交叉熵做训练损失函数,当损失值变小时评价指标不一定变得更优。为解决上述缺陷,提出一种基于度量指标损失的U-Net网络模型。由于错误接受率和错误拒绝率变小时度量指标F-Mea-sure会上升,因此构建半错误率损失函数。采用分治策略,将文档图像分割成固定大小的图像块,分别进行二值化。在文档图像竞赛数据集上进行大量对比实验,实验结果表明,该方法相比原始U-Net,在4个度量指标上均有提升,二值化结果图像的文字连通性更好、噪声更少。At present,the deep neural networks generally use cross-entropy as the training loss function.When the loss score becomes small,the evaluation measure index may not become better.To solve the above defect,a U-Net network based on measure index loss was proposed.When the false acceptance rate and false rejection rate become smaller,the F-Measure increases.Therefore,a measure index loss function named half total error rate was constructed.The divide-and-conquer strategy was adopted to divide the document image into many image patches with fixed size,each patch was binarized separately.Extensive experiments were conducted on eight document image competition datasets.Experimental results show that this method gets better results for the four metrics when compared with the original U-Net.Moreover,the binarized image has better text connectivity and less noise.

关 键 词：文档图像二值化 卷积神经网络 交叉熵 度量指标损失 打印图像 手写图像 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]