基于相对生成对抗网络的低清小脸幻构  

Tiny Face Hallucination via Relativistic Adversarial Learning

作　　者：邵文泽[1,2] 张苗苗 李海波 SHAO Wenze;ZHANG Miaomiao;LI Haibo(College of Telecommunications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;Key Laboratory of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京邮电大学通信与信息工程学院,南京210003 [2]南京理工大学高维信息智能感知与系统教育部重点实验室,南京210094

出　　处：《电子与信息学报》2021年第9期2577-2585,共9页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(61771250,61972213,11901299);中央高校基本科研业务费专项资金(30918014108)。

摘　　要：针对当前代表性低清小脸幻构方法存在的视觉真实感弱、网络结构复杂等问题,提出了一种基于相对生成对抗网络的低清小脸幻构方法(tfh-RGAN)。该文方法的网络架构包括幻构生成器和判别器两个部分,通过像素损失函数和相对生成对抗损失函数的联合最小化,实现生成器和判别器的交替迭代训练。其中,幻构生成器结合了残差块、稠密块以及深度可分离卷积算子,保证幻构效果和网络深度的同时降低生成器的参数量;判别器采用图像分类问题中的全卷积网络,通过先后去除批归一化层、添加全连接层,充分挖掘相对生成对抗网络在低清小脸幻构问题上的能力极限。实验结果表明,在不额外显式引入任何人脸结构先验的条件下,该文方法能够以更简练的网络架构输出清晰度更高、真实感更强的幻构人脸。从定量角度看,该文方法的峰值信噪比相较之前的若干代表性方法可提高0.25~1.51 dB。Considering that previous tiny face hallucination methods either produced visually less pleasant faces or required architecturally more complex networks,this paper advocates a new deep model for tiny face hallucination by borrowing the idea of Relativistic Generative Adversarial Network(tfh-RGAN).Specifically,a hallucination generator and a relativistic discriminator are jointly learned in an alternately iterative training fashion by minimizing the combined pixel loss and relativistic generative adversarial loss.As for the generator,it is mainly structured as concatenation of a few basic modules followed by three 2×up-sampling layers,and each basic module is formulated by coupling the residual blocks,dense blocks,and depthwise separable convolution operators.As such,the generator can be made lightweight while with a considerable depth so as to achieve high quality face hallucination.As for the discriminator,it makes use of VGG128 while removing all its batch normalization layers and embedding a fully connected layer additionally so as to fulfill the capacity limit of relativistic adversarial learning.Experimental results reveal that,the proposed method,though simpler in the network architecture without a need of explicitly imposing any face structural prior,is able to produce better hallucination faces with higher definition and stronger reality.In terms of the quantitative assessment,the peak signal-to-noise ratio of the proposed method can be improved up to 0.25~1.51 dB compared against several previous approaches.

关 键 词：图像处理 超分辨率 人脸幻构 深度学习 生成对抗网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]