Document Image Binarization with Fully Convolutional Neural Networks 图片文档二值化FCNs

FCNs

1.abstract
提出了FCNN，The FCN is trained to optimize a continuous version of the Pseudo F-measure metric and an ensemble of FCNs outperform the competition winners on 4 of 7 DIBCO competitions.
指出这个模型在Palm Leaf Manuscripts里面效果也很好

2.introduction
指出在识别文档之前应该先二值化
我们提出的FCNN能对各种文档图片作二值化并且不需要调参
回顾其它传统方法，指出FCNs learn from training data to exploit the spatial arrangements of pixels without relying on a handcrafted bias on local shapes.
本文贡献：
1.提出FCNs和架构在二值化上的使用
2.We show that directly optimizing the proposed continuous Pseudo F-measure exceeds the previous state-of-the-art on DIBCO competition data.
3.通过计算学习曲线得出结论：数据多样性比数据质量重要
4.证明了在数据输入特征更多的时候，FCNN表现更出色

3.related work
总结了前人的方法
4.methods
A.FULLY CONVOLUNTIONAL NETWORKS
介绍了结构，各个输入输出尺寸和relu函数
B.MULTI-SCALE
提出不同尺寸输入合并能提高性能效果
介绍了上采样和下采样结构
C.Pseudo F-measure Loss
看不懂
D.Datasets and Metrics
用了两个数据集 DIBCOs [1]–[7] and Palm Leaf Manuscripts (PLM)
E.Implementation Details
看不懂，讲了下训练细节

5.experiments
A.loss functions
用了四个损失函数：P-FM, FM, P-FM + FM, and Cross Entropy (CE)
其中，由于P-FM有点问题（predicting border pixels as background）？就用了P-FM+FM一起，CE loss is a standard classification based loss
B.DIBCO performance
将自己FCNs与冠军队伍作了对比
C.Architecture Search
作者更改了一些结构的超参数（depth,width,kernel,scale），发现效果提升不大，认为对于数据集的更改比结构更改更好
D.How Much Data is Enough?
数据增多可能会提高效果，但太多会适得其反，绘制了关于数据量的学习曲线
提出多样性的training集至关重要
E. Input Features
指出Relative Darkness特征的重要性
We used RD features with a window size of 5x5 and a similarity threshold of ±10 in all experiments in this paper.

6.conclusion
首先，我们结合P-FM和FM损失函数，并用FCNs训练文档图片二值化
其次，指出训练集的多样性比其它重要很多
最后，we analyzed using additional features as input to the FCN and found that Relative Darkness features [26] and the output of Howe binarization [9] perform best.