OHW系列之一《Stroke Sequence-Dependent DCNN for Online Handwritten Chinese Character Recognition 》
一、论文摘要
本文提出一种名为Stroke Sequence-dependent Deep Convolutional Neural Network(SSDCNN)的模型，模型将Stroke sequence information 和八方向特征结合，用于在线手写中文字符识别。
首先，SSDCNN将Stroke Sequence依据其书写顺序转换为堆栈的特征图作为模型输入，接着经过深度卷积网络提取特征与八方向特征结合作为手写轨迹的最终特征表示。
模型训练策略：“we design a two stages algorithm to train SSDCNN. Firstly, we use
the data to pre-train the whole framework. In order to merge the two representation efficiently, we further train the fullly-connected multiple layer perceptron with a softmax
layer by fixing the parameter of the DCNN.”
二、论文的国内外现状
本文主要是解决在线手写的单字符识别，论文国内外现状从传统方法和现代方法进行了分析。
传统方法流程：
1. 数据预处理
- Data Normalization: 将Strokes正则化到一个固定的大小中以便特征提取，但如何恢复手写字符的外型的形变是关键！常用的方法如：Line Density-based Normalization, Curve-fitting-based normalization and Pseudo 2D Normalization. (“The NLN based on line density equalization has been proven very efficient [16]. In
order to use line density-based method to OLHCC, Phan[17] convert an OLHCC to a 2D image. For Curve-fitting-based normalization methods, three algorithm are often used such as bi-moment normalization(BMN) [18], centroid-boundary alignment (CBA), and modified CBA (MCBA) [19]. They can be used to OLHCC after
slightly modified [2, 17]. Pseudo two-dimensional normalization is popularly used for off-line HCCR[15]”). - Interpolation: Linear Interpolation, Bezier Interpolation, B-spline Interpolation.
2. 特征提取
- Directional features
- Gradient features
三、论文算法模型
模型概述：
- The DCNN is used to learn the high-quality representation of OLHCCs, which is stroke sequence-dependent and can learn the stroke sequence information and structural shape of OLHCCs from large scale instances.
- The statistical feature is integrated into SSDCNN via a deep fully-connected neural network. Specifically, the eight-directional features are used, which has been proven efficient in OLHCCR. Hence, SSDCNN can preserve the strength of statistical features.
模型细节：
上分支：模型依据单个字的笔画，将每一笔笔画作为输入，对于不同的字笔画数量一般不同则pad到该batch中最多的笔画的数量。将所有的笔画作为输入送到深度神经网络中提取特征图。
下分支：下面的分支采用传统的方法进行八方向特征提取，八方向特征通过数学统计的方法分析该点及周围点的来表示笔画的形状及方向，缺点在于它不能体现除了该笔画以外的信息，“, because it only extracts the direction feature between the ending point of one stroke and the starting point in its following stroke regardless of the positional relationship between other points in different strokes.”
对于两个分支分别输入到不同的全链接层，原因在于两种特征表示图分别采用不同的方法来提取，卷积网络中经过relu函数**后，特征中数值从0-inf, 而八方向特征中数值范围仍然为[0,1]。
四、论文实验结果
模型的实验结果感觉意义不大，主要是做的单文字识别。
五、论文创新点
模型主要提供了一种思路如何结合传统的特征提取方法，如：八方向特征、signature of path, 与现有的深度卷积神经网络结合，但是该文的局限在于仅对单字符识别而且将字符拆解为一个个固定size的笔画作为输入，数据冗余较大。