一种基于生成式对抗网络的图像数据扩充方法  

An Image Data Augmentation Method Based on Generative Adversarial Network

作　　者：王海文 邱晓晖[1] WANG Hai-wen;QIU Xiao-hui(School of Telecommunications&Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003

出　　处：《计算机技术与发展》2020年第3期51-56,共6页Computer Technology and Development

基　　金：江苏省自然科学基金(BK2011789)。

摘　　要：针对卷积神经网络(CNN)在数据集(训练集)较小时,易发生过度拟合的现象,提出并实现了一种引入Selu激活函数并结合带参数归一化的Dropout方法的深度卷积生成式对抗网络用于图像增强,生成图像实现数据集扩充,从而解决深度学习图像分类研究中因图像数据不足造成的模型表达能力差、训练时易过度拟合的问题。通过裁剪、旋转、插值、畸变变换等扩充图像集的传统图像增强方法往往只能扩充样式单一甚至信噪比较低的图像,与传统图像增强方法扩充图像集不同,使用生成式对抗网络生成的图像明显区别于原始图像,不仅可以得到数量更多,内容更丰富的高质量图像,数据集扩充效率也得以提升。仿真实验表明,该生成式对抗网络得到了质量相对较高的图像,有效地扩充了数据集。Aiming at the problem of over-fitting caused by too small datasets in convolution neural networks(CNN),we propose and implement a deep convolution generative adversarial neural network with Selu activation function and Dropout method with parameter normalization to achieve image expansion through image generation and image augmentation,so as to solve the problems of poor model expression ability and easy over-fitting in training caused by insufficient image data in deep learning image classification research.Traditional image augmentation methods of expanding datasets by cropping,rotating,interpolating and distortion transform can only expand the image with a single content or even low SNR.Different from the traditional image augmentation method to expand the datasets,the image generated by the generative adversarial network is obviously different from the original image,which can not only obtain more high-quality images with richer content,but also improve the efficiency of dataset expansion.The simulation experiment shows that the DCGAN with Selu activator can effectively make a high-quality image generation that finally enlarge the image dataset.

关 键 词：深度卷积 生成式网络 图像增强 数据增强 

分 类 号：TN911.23[电子电信—通信与信息系统]