利用VGGnet对印章印文分类识别的适用条件研究

Applicable Conditions for Classification and Recognition of Seals and Stamps by Using VGGnet

为探究利用卷积神经网络VGGnet对印章印文分类识别的适用条件,利用VGGnet对固态光敏印章、塑胶印章、铜章,共三枚印章盖印的21 000枚印文样本进行印章印文分类识别,通过改变样本量大小、迭代次数、学习率等,探究实验的适用条件。实验发现以18 000枚印文作为训练样本、迭代学习10次、在0.001的学习率下进行特征学习,训练得到的VGGnet模型对300枚印文进行分类识别准确率达到100%,损失值降低为0.000 2。结果表明:VGGnet可以作为一种辅助方法应用于印章印文自动识别中,实验结果为以后的研究提供了参考依据。

In order to explore the applicable conditions for classification and identification of seals and stamps by using the convolutional neural network VGGnet,the VGGnet is used to classify and identify seals and stamps of 21 000 printed samples of the solid photosensitive seal,plastic seal and bronze seal.Methods such as changing the sample size,number of iterations,and learn-rate are used to explore the applicable conditions of the experiment.The experiment indicates that with the 18 000 pieces of printed text as training samples,and via 10 times of iterative learning and feature learning at learning rate of 0.001,the VGGnet model obtained by training could reach a classification and recognition accuracy of 100%on 300 stamps,and the loss value is reduced to 0.000 2.The experiment results show that VGGnet may be used as an auxiliary method in the automatic identification of seal printing,thus providing a reference for future research.