基于多尺度特征融合的属性已知人脸图像超分辨率  

Multi-scale feature fusion based attribute-aware face image super-resolution

作　　者：姚肇亮 刘宇男 张姗姗 杨健[1] Yao Zhaoliang;Liu Yunan;Zhang Shanshan;Yang Jian(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学计算机科学与工程学院,江苏南京210094

出　　处：《南京理工大学学报》2022年第5期571-578,共8页Journal of Nanjing University of Science and Technology

基　　金：中央高校基本科研业务费专项资金资助(30920032201);国家自然科学基金(62172225)。

摘　　要：卷积神经网络凭借其强大的表征能力,在图像超分辨率任务上取得了许多令人满意的结果。许多基于神经网络的方法采用增加网络深度的方式,存在存储空间消耗多、实用性不强的问题。为解决该问题,该文提出一种基于多尺度特征融合的属性感知人脸图像超分辨率网络。该文借助局部残差模块和逐元素相加的融合方式以减少网络复杂性并提炼出表征能力优秀的多尺度特征。该文构建一个可自适应地融合多尺度特征和人脸先验的属性感知模块,使得网络学习到更丰富的语义信息。该文提出的网络由多个网络子模块级联构成,并通过一个多层次特征融合模块进行共同学习。试验表明:该文方法能取得良好的超分辨率性能,输出更加真实的人脸图像,可以通过调整人脸属性信息进行人脸图像生成效果的操纵。Convolutional neural networks(CNN)have achieved many satisfactory results in face image super-resolution tasks with its powerful representation ability.However,most of the CNN-based methods adopt an increasing network depth approach which demands more storage space and are less applicable in practice.In order to solve this problem,this paper proposes a multi-scale feature fusion based attribute-aware face image super-resolution network.First,this paper adopts a local residual learning via shortcut connection and element-wise addition to reduce the complexity and distill the representative multi-scale features.Second,this paper adopts an attribute-aware module to adaptively rescale and fuse multi-scale features and facial prior which provide rich semantic information to help better super-resolve face images.Third,the network proposed here is constructed by a set of sub-modules,in which the features from each sub-module can be jointly learned by using a hierarchical feature fusion module.Extensive experiments show that the proposed network achieves more appealing results than the state-of-the-art methods;moreover,the method not only generates perceptually plausible face images,but also is able to manipulate the results by adjusting the attribute information.

关 键 词：人脸图像超分辨率 多尺度特征融合 人脸先验信息 深度卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]