摘要
自然图像增强是计算机视觉领域中的一个研究热点.针对以往图像增强方法计算过程复杂和参数需手工设置等缺陷,提出一种基于条件生成对抗模型的图像增强(enhancement with conditional generative adversarial networks,E-CGAN)方法.分别构建生成式神经网络和判别式神经网络,其中,生成模型直接对图像进行处理生成最终增强的图片结果,判别模型在训练阶段对生成模型构建对抗型损失函数,优化生成模型的参数.在生成模型的结构中,加入L1距离误差函数作为生成模型的约束,并提出连续多尺度跨层连接方式,加快网络的训练速度,提高生成模型的准确率.在图像清晰度增强,灰度图像着色两个图像增强问题上进行实验,结果表明,E-CGAN可以有效地保留图像特征,PSNR和SSIM质量平均提高9%和5%.
Wild image enhancement technology is a hotspot in the field of computer vision.To overcome the defects on complexity of calculation and manual setting parameters of the conventional image enhancement methods,a novel image enhancement method with conditional generative adversarial network(E-CGAN)has been proposed.The generative neural network and the discriminant neural network are constructed respectively,where the generative model is used to generate the final images and the discriminant model is employed to construct the confrontation loss function in the training stage,so as to optimize the parameters of the two models.In the structure of the generative model,a successive-multiple skip connection method constrained by L1 error function is proposed,which speeds up the training speed of the network and improves the accuracy of the generative model.Two implementations on image sharpening and colorization have been implemented to evaluate the effectiveness of the proposed method,the experimental results show that E-CGAN can effectively highlight the characteristics of the image,and better quality promotion achieves up to 9%and 5%both on PSNR and SSIM index.
作者
贾玉福
胡胜红
刘文平
王超
向书成
Jia Yufu;Hu Shenghong;Liu Wenping;Wang Chao;Xiang Shucheng(Information Management and Statistics School,Hubei University of Economics,Wuhan 430205,China;Information and Communication Engineering School,Hubei University of Economics,Wuhan 430205,China)
出处
《南京师大学报（自然科学版）》
CAS
CSCD
北大核心
2019年第3期88-95,共8页
Journal of Nanjing Normal University(Natural Science Edition)
基金
国家自然科学基金(61572012)
教育部人文社科项目(18YJCZH050)
湖北省自然科学基金(D20182202)
教育厅科研计划(2018CFB721)
关键词
图像增强
生成对抗网络
深度学习
图像质量
image enhancement
generative adversarial network
deep learning
image quality
