基于生成对抗网络的多用途图像增强鲁棒算法  

MULTIPURPOSE IMAGE ENHANCEMENT ROBUST ALGORITHM BASED ON GENERATIVE ADVERSARIAL NETWORK

作　　者：李英[1] 李臻[2] Li Ying;Li Zhen(Neijiang Vocational and Technical College,Neijiang 641000,Sichuan,China;University of Electronic Science and Technology of China,Chengdu 610054,Sichuan,China)

机构地区：[1]内江职业技术学院信息技术系,四川内江641000 [2]电子科技大学,四川成都610054

出　　处：《计算机应用与软件》2020年第6期247-252,262,共7页Computer Applications and Software

基　　金：四川省重点实验室开放基金项目(GK201608);电子科技大学中山学院校内项目(418YKQN04)。

摘　　要：针对现有图像增强技术不能同时进行超分辨率和伪影去除的问题,提出一种基于生成对抗网络的图像增强鲁棒算法,能够在一个网络中以端到端的方式同时进行伪影消除和超分辨率。算法包括生成器网络和判别器网络两部分。生成器网络以U-net形式引入跳跃链接进行共享信息,并在最后一层之前设计一个像素解析模块,提高输出图像分辨率;判别器采用自动编码方式有助于将大量有关生成图像质量的语义信息传递回生成器;提出一种基于网络特征损失、边缘损失和判别器重构损失三者加权的感知损失函数,有效保留在图像增强过程中经常丢失的锐度。实验结果表明,对于高压缩低分辨率图像,该方法可以同时进行伪影去除和超分辨率,相对其他方法在多个评价指标上都有很好的性能体现。Aiming at the problem that the existing image enhancement techniques cannot simultaneously carry on super-resolution and artifacts removal,this paper proposes an image enhancement robust algorithm based on the generative adversarial network.It can eliminate artifacts and super-resolution stimultaneously in an end-to-end way in a network.The algorithm includes two parts:generator network and discriminator network.The generator network introduced skip links in the form of U-Net to share information.A pixel resolution module was designed before the last layer to improve the resolution of the output image.The autoencoder of the discriminator was helpful to transfer a lot of semantic information about the quality of the generated image back to the generator.We proposed a weighted perceptual loss function based on network feature loss,edge loss and discriminator reconstruction loss,and it could effectively preserve the sharpness that was often lost in the process of image enhancement.The experimental results show that our method can perform artifacts removal and super-resolution simultaneously for high compression and low-resolution images.Compared with other methods,our method has good performance in many evaluation indicators.

关 键 词：生成对抗网络 超分辨率 伪影去除 图像增强 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]