«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

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

卷:

31

期数:

2021年11期

页码:

95-100

栏目:

图形与图像

出版日期:

2021-11-10

文章信息/Info

Title:

Low Illumination Face Image Enhancement Based on CycleGAN

文章编号:

1673-629X(2021)11-0095-06

作者:

朱克亮¹ ; 张天忠² ; 石雪梅¹ ; 张树涛³; 4 ; 陈良锋³

1. 国网安徽省电力有限公司建设分公司,安徽 合肥 230022;

2. 国网安徽省电力有限公司,安徽 合肥 230061;

3. 中国科学院合肥物质科学研究院,安徽 合肥 230026;

4. 中国科学技术大学,安徽 合肥 230031

Author(s):

ZHU Ke-liang¹ ; ZHANG Tian-zhong² ; SHI Xue-mei¹ ; ZHANG Shu-tao³; 4 ; CHEN Liang-feng³

1. State Grid Anhui Electric Power Co, Ltd. Construction Company,Hefei 230022,China;

2. State Grid Anhui Electric Power Co, Ltd. ,Hefei 230061,China;

3. Hefei Institutes of Physical Science,Chinese Academy of Sciences,Hefei 230026,China;

4. University of Science and Technology of China,Hefei 230031,China

关键词:

循环生成对抗网络; 图像增强; 低照度; 梯度处罚; 深度学习

Keywords:

cycle generative adversarial networks; image enhancement; low illumination; gradient penalty; deep learning

分类号:

TP311. 5

DOI:

10. 3969 / j. issn. 1673-629X. 2021. 11. 016

摘要:

人脸识别系统经常会受到光照环境的影响。 为了提高低照度条件下的人脸识别性能,提出了一种基于循环生成对抗网络的低照度人脸图像增强方法,利用循环生成对抗网络将低照度条件下的人脸图像转换成正常光照下的人脸图像。 模型包含生成器和判别器, 生成器由包含 4 个卷积层、9 个残差网络层和 2 个转置卷积层的卷积神经网络组成,判别器由包含 5 个卷积层的卷积神经网络组成。 在循环生成对抗网络训练的过程中,采用改进的损失函数,并结合梯度惩罚项来训练网络模型,提升了稳定性,加快了网络收敛速度,并且提高了生成人脸图像的质量。 在 VV 数据集上的实验结果表明, 该方法能够有效地实现低照度条件下的人脸图像增强,对比 HE、MSR 和 MSRCR 算法在 PSNR、SSIM 和 MSE 指标上有较大提高。

Abstract:

Face recognition systems are often affected by the lighting environment. To improve the performance of face recognition under low illumination conditions, a low illumination face image enhancement method based on cycle - consistent adversarial network is proposed. The network that consists of generators and discriminators is used to transfer low illumination face images to normal illumination face images. The generators consist of convolutional neural networks with four convolution layers,nine residual network layers and two transposed convolution layers,and? ? ?the discriminators consist of convolutional neural networks with five convolution layers.In the training process of the cycle-consistent adversarial network,the improved loss function and the gradient penalty are used to train the network,which improves the stability,speeds up the network convergence and improves the quality of the generated face image. Experimental results on VV dataset show that the proposed method can effectively achieve face image enhancement under low illumination conditions,and compared with HE,MSR and MSRCR,it improves PSNR,SSIM and MSE values.

常用功能

更新日期/Last Update: 2021-11-10