基于高效注意力机制的人脸超分辨网络  

Face Super-Resolution Network Based on Efficient Attention Mechanisms

作　　者：许子祥 豆锐 李佳雯 高广谓 XU Zixiang;DOU Rui;LI Jiawen;GAO Guangwei(School of Automation,Nanjing University of Posts and Telecommunications,Nanjing 210023;Provincial Key Laboratory for Computer Information Processing Technology,Soochow University,Suzhou 215000)

机构地区：[1]南京邮电大学自动化学院,南京210023 [2]苏州大学江苏省计算机信息处理技术重点实验室,苏州215000

出　　处：《计算机与数字工程》2024年第10期3065-3071,共7页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:61972212);江苏省自然科学基金项目(编号:BK20190089);江苏省“六大人才高峰”项目(编号:RJFW-011);苏州大学江苏省计算机信息处理技术重点实验室开放课题(编号:KJS1840)资助。

摘　　要：目前,通过与面部先验联合的深度学习方法已经能够较好地恢复退化后的人脸图像细节。然而这些方法存在着一些不足。一方面,多任务联合训练需要在数据集上进行额外先验标记,并且引入先验预估将显著增加网络的计算量。另一方面,大多数方法仅使用卷积神经网络(CNN),然而CNN有限的感受野会降低重构的完整性和面部图像的准确性。针对上述问题,提出了一种基于高效注意力机制的人脸超分辨率网络。该网络在不需要任何先验信息的辅助下,通过结合CNN与Transformer等注意力机制,能够有效地恢复人脸的全局结构与局部纹理细节。在各个数据集的大量实验都表明该网络可以达到理想的结果。Currently,deep learning methods based on facial priors have been able to better recover the details of degraded face images.However,these methods still have some shortcomings.On the one hand,multi-task joint training requires additional prior labeling on the dataset,and the introduction of the priors will significantly increase the computational cots of the network.On the other hand,most methods only use convolutional neural networks(CNN),the limited perceptive field of CNN will reduce the in-tegrity and accuracy of the reconstructed facial images.To address these problems,a face super-resolution network based on effi-cient attention mechanisms is proposed.The network combines CNN with several attention mechanisms such as Transformer.It can effectively recover the global structure and local texture details of a face without the aid of any facial priors.Extensive experiments on various datasets show the effectiveness of the proposed network.

关 键 词：人脸超分辨率 TRANSFORMER 卷积神经网络 注意力机制 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]