融合边缘注意力的手写蒙古文字元数据增强方法  

Feature-Level Handwritten Mongolian Text Metadata Enhancement Method Incorporating Edge Attention

作　　者：石佳钰 殷雁君 张文轩 智敏 SHI Jia-yu;YIN Yan-jun;ZHANG Wen-xuan;ZHI Min(College of Computer Science and Technology,Inner Mongolia Normal University,Hohhot 010022,China)

机构地区：[1]内蒙古师范大学计算机科学技术学院,内蒙古呼和浩特010022

出　　处：《内蒙古师范大学学报（自然科学汉文版）》2023年第2期189-196,共8页Journal of Inner Mongolia Normal University(Natural Science Edition)

基　　金：内蒙古自治区自然科学基金资助项目(2021LHMS06009)。

摘　　要：针对手写蒙古文字元数据集样本少且多样性差的问题,提出融合边缘注意力的条件手写蒙古文字元生成模型。模型在条件生成对抗网络的基础上引入了边缘注意力机制,使得数据生成模块对手写蒙古文字元边缘变化更加敏感,增加特征多样正则项在一定程度损失避免模式崩溃并使得生成样本更具多样性。在MNIST和手写蒙古文数据集进行了大量实验,结果表明提出模型样本增强效果优于GAN、CGAN,并且增强后的样本能够有效提升文字识别模型的性能。In the light of the problem of less samples and poor diversity in the handwritten Mongolian metadata set, a conditional generation model of handwritten Mongolian characters with edge attention is proposed in the paper. The model introduces an edge attention mechanism on the basis of conditional generative adversarial network, which makes the data generation module more sensitive to the edge changes of handwritten Mongolian characters, and adds a regular term with diverse features to a certain extent to avoid mode collapse and make the generated samples more diverse. A large number of experiments are conducted on the MNIST and handwritten Mongolian datasets, which show that the sample enhancement effect of the proposed model is better than that of GAN and CGAN, and the enhanced samples are capable of effectively improving the performance of the text recognition model.

关 键 词：数据增强 CGAN 手写蒙古文字元 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]