基于非参数回归和卷积神经网络的在线手写签名身份认证模型研究

Online Handwriting Signature Authentication Model Study Based on Nonparametric Regression Combined with CNN

在线手写签名身份识别模型受训练签名的变异复杂性、签名数量有限、真伪签名不均、动态特征等实际因素影响,需要对签名轨迹数据曲线拟合与深度学习的图像特征提取的联合建模.文章设计了在三种签名变异复杂性情形下,比较了不同的非参数回归和卷积神经网络建模的试验效果.实验的主要发现如下：1.单独使用CNN可获得较高的签名鉴别率,其效果好于单独使用非参数回归的效果,代价是需用较为复杂的神经网络结构,训练时间可能过长;2.真伪签名轨迹样本方差和训练样本量不均衡都会不同程度地影响到CNN的学习效果;3.签名的动态特征相较于仅使用静态特征而言有助于提升模型的识别能力;4.当真假签名数量有限,只用样条回归或CNN的独立判定效果都不甚理想,选择合适自由度的样条回归与CNN联合建模在达到可用精度条件下,可有效降低训练时间,降低卷积神经网络的核个数,减小模型复杂性,达到较好的识别性能.

Online handwritten signature authentication Modeling includes the dynamic characteristics,the variability of the training signature, the limited number of signatures, unbalanced size of actual signature and forged signature. It is necessary to explore joint modeling of image feature extraction for signature track curve fitting and deep learning. In this paper, experimental results of different nonparar metric regression and Convolutional Neural Network（CNN） algorithms are explored under three kinds of signature variations. The main discoveries are as follows： 1. The performance when just using CNN can obtain higher identification than only using nonparametric regression. However, training cost is paid by more complicated neural network structure and long training time is consumed; 2. Signature sample trajectory variance and unbalanced sample size will influence to CNN learning effect; 3. Dynamic characteristics of the signature is helpful to enhance the recognition compared with only use static characteristics; 4. Under limited number of sample, solution with only spline regression or CNN is not ideal. By selecting the appropriate degrees of freedom of spline regression and parameters in CNN can effectively reduce the training time, reduce the number of nuclear of convolution neural network which reduce model complexity and achieve desirable recognition performance.