基于MUNIT的数据增强方法研究  

Study of Data Augmentation Method Based on MUNIT

作　　者：丁柏楠 王淏 袁文翠[1] 吴圣洁 DING Bonan;WANG Hao;YUAN Wencui;WU Shengjie(School of Computer & Information Technology, Northeast Petroleum University, Daqing 163319, China;College of Physics, Jilin University, Changchun 130012, China)

机构地区：[1]东北石油大学计算机与信息技术学院,黑龙江大庆163319 [2]吉林大学物理学院,吉林长春130012

出　　处：《微型电脑应用》2021年第12期19-22,共4页Microcomputer Applications

基　　金：黑龙江省创新创业训练计划项目(202010220083);黑龙江省教育厅科学技术研究项目(12521050)。

摘　　要：近年来深度学习和机器学习在图像分类上取得了重大的进展,但高效的网络训练需要巨量的数据支持。在只有少量数据的情况下,模型的表现十分糟糕。数据增强方法可以通过更加高效的使用已有的数据来减轻该问题。但通常的数据增强方法只通过裁剪,旋转等方法产生有限的可能有用的数据。针对这一问题,提出了一种基于MUNIT的数据增强方法,通过使用MUNIT基于已有的少量数据来生成新的数据。据实验结果表明,该方法可以生成具有相同内容但不同类别的新数据,将新的数据加入神经网络训练后,神经网络不会出现过拟合的情况而且分类的准确度有了一定程度上的提高。通过与传统的数据增强方法相比,该方法得到的结果更好,以此验证了该方法的可行性。In recent years,deep learning and machine learning have made significant progress in image classification,but efficient network training requires huge amounts of data support.With only a small amount of data,the model performed very badly.Data augmentation methods can alleviate this problem by using existing data more efficiently.However,the usual data augmentation methods only produce limited possible useful data through methods such as cropping and rotation.In response to this problem,this paper proposes a data augmentation method based on MUNIT,to generates new data based on a small amount of existing data by using MUNIT.According to the experimental results,this method can generate new data with the same content but different categories.After adding the new data to the neural network training,the neural network will not be over-fitted and the classification accuracy has been improved to a certain extent.Compared with traditional data augmentation methods,the proposed method obtains better results,which verifies the feasibility of the method.

关 键 词：生成对抗网络 数据增强 图片分类 卷积神经网络 深度学习 

分 类 号：TN911.73[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]