[1]黄志伦,刘 俊,郑 萌.改进 GAN 的光化性角化病图像数据增强方法[J].计算机技术与发展,2022,32(09):36-42.[doi:10. 3969 / j. issn. 1673-629X. 2022. 09. 006]
　HUANG Zhi-lun,LIU Jun,ZHENG Meng.An Improved Generation Adversarial Networks Method for Data Augmentation of Actinic Keratosis Image[J].,2022,32(09):36-42.[doi:10. 3969 / j. issn. 1673-629X. 2022. 09. 006]

点击复制

改进 GAN 的光化性角化病图像数据增强方法()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

32

期数:

2022年09期

页码:

36-42

栏目:

媒体计算

出版日期:

2022-09-10

文章信息/Info

Title:

An Improved Generation Adversarial Networks Method for Data Augmentation of Actinic Keratosis Image

文章编号:

1673-629X(2022)09-0036-07

作者:

黄志伦¹; 2 ; 刘 俊¹; 2 ; 郑 萌³

1. 武汉科技大学 计算机科学与技术学院,湖北 武汉 430065

2. 智能信息处理与实时工业系统湖北省重点实验室(武汉科技大学),湖北 武汉 430065

3. 武汉理工大学 计算机科学与技术学院,湖北 武汉 430070

Author(s):

HUANG Zhi-lun1; 2 ; LIU Jun1; 2 ; ZHENG Meng3

1. School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China

2. Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System ( Wuhan University of Science and Technology) ,Wuhan 430065,China

3. School of Computer Science and Technology,Wuhan University of Technology,Wuhan 430070,China

关键词:

光化性角化病; 生成对抗网络; 数据增强; 高斯混合模型; U-Net 变体; 混合体驱动

Keywords:

actinic keratosis;  generative adversarial networks;  data augmentation;  Gaussian mixed model;  modification on U - Net; mixture boost

分类号:

TP391

DOI:

10. 3969 / j. issn. 1673-629X. 2022. 09. 006

摘要:

针对传统网络应用于医学图像数据增强时出现的特征丢失和多样性不足的问题,提出一种基于能量的改进生成对抗网络模型。 首先将简单的原始随机噪声输入高斯混合模型,尽管增加了少部分的计算量,但转换后具有潜在复杂分布的噪声能在一定程度上提升生成样本的类内多样性和类间多样性。 然后在判别器部分把简单自编码器替换为 U-Net状的变体网络,多层采样的过程可以增强对细节纹理的感知,进而提高生成图像的清晰度和特征还原。 最后使用混合体驱动算法,按照加权参数逐步混合多个生成器,在迭代过程中弥补先前混合体的不足,提高生成模块的鲁棒性。 在光化性角化病图像数据集上的实验结果表明,该网络训练生成的图像在弗雷歇初始距离上优于现有的 WGAN( Wasserstein GAN)模型 3. 41。 另外由于判别器可预训练的特性,收敛速度快于当前的 WGAN。 同时也在公开数据集 MNIST 和 CelebA 上验证改进生成对抗网络的有效性。

Abstract:

Aiming at the problems of feature loss and lack of diversity when traditional network is applied to medical image data enhancement,an improved generation countermeasure network model based on energy is proposed. Firstly,the simple original random noise is input into the Gaussian mixture model. Although a small amount of calculation is increased,the converted noise with potentially complex distribution can improve the intra class diversity and inter class diversity of the generated samples to a certain extent. Then in the discriminator part,the simple self encoder is replaced by a U-Net like variant network. The multi-layer sampling process can enhance the perception of detail texture, and then improve the clarity and feature restoration of the generated image. Finally, the hybrid driven algorithm is used to gradually mix multiple generators according to the weighted parameters to make up for the shortcomings of the previous hybrid in the iterative process and improve the robustness of the generation module. The experimental results on the actinic keratosis image data set show that the image generated by the network training is better than the existing WGAN ( Wasserstein GAN) model 3. 41 in the FID score. In addition,because the discriminator can be pre trained,the convergence speed is faster than the currentWGAN. At the same time,the effectiveness of the improved generation countermeasure network is also verified on the public datasets MNIST and CelebA.

常用功能