基于特征选择和幻象残差网络的在线手写签名认证
Online Handwritten Signature Verification Based on Feature Selection and Ghost Residual Network

为了解决当前在线手写签名认证(online signature verification, OSV)模型无法从有限的签名样本中提取稳定的签名特征以及模型参数量大问题，本文提出了基于特征选择和一维幻象残差网络(1D-GRNet)的在线手写签名认证方法。首先，采用随机森林特征选择算法对各个签名的全局特征集进行筛选，获得足够稳定的签名特征，以此来提高签名认证的准确率；然后，采用幻象模块对标准残差结构进行改进，构建一维幻象残差模块，降低整体网络模型的参数量，同时为签名小样本训练提供可能，提高了模型的实用性。最后，本文方法在数据集MYCT-DB1和SVC2004-task2上进行验证。当采用5个真伪签名进行小样本进行训练时，在两个数据集上的等错误率分别为3.21%和4.57%。当采用10个真伪签名进行训练时，在两个数据集上的等错误率分别为1.53%和2.93%。实验结果表明所提方法能够有效提高签名认证精度。
In order to solve the problem that online signature verification (OSV) model cannot extract stable signature features from limited signature samples and have a large number of model parameters, this paper proposes an online handwritten signature verification method based on feature selection and one-dimensional ghost residual network (1D-GRNet). Firstly, the random forest feature selection algorithm is adopted to screen the global feature set of each signature to obtain sufficiently stable signature features, so as to improve the accuracy of signature verification. Then, the ghost module is adopted to improve the standard residual structure, and a one-dimensional phantom residual module is constructed to reduce the number of parameters of the overall network model, and at the same time provide the possibility for signature small sample training, which improves the practicability of the model. Finally, the method is verified on the datasets MYCT-DB1 and SVC2004-task2. When 5 genuine and forged signatures are adopted for small sample training, the equal error rates (EERs) on the two datasets are 3.21% and 4.57%, respectively. When 10 genuine and forged signatures are adopted for training, the EERs on the two datasets are 1.53% and 2.93%, respectively. The experimental results represent that the proposed method can effectively improve the accuracy of signature verification.