基于Transformer与注意力聚合的人脸超分辨率  

Face Super-Resolution Network Based on Transformer and Attention Integration

作　　者：唐雷 许子祥 高广谓 TANG Lei;XU Zixiang;GAO Guangwei(College of Automation&College of Artificial Intelligence,Nanjing University of Posts and Telecommunications,Nanjing 210023;Institute of Advanced Technology,Nanjing University of Posts and Telecommunications,Nanjing 210023)

机构地区：[1]南京邮电大学自动化学院人工智能学院,南京210023 [2]南京邮电大学先进技术研究院,南京210023

出　　处：《计算机与数字工程》2023年第12期2977-2983,共7页Computer & Digital Engineering

摘　　要：目前大多数人脸超分辨率方法都是将整个人脸图像或利用额外的人脸先验输入到卷积神经网络中来辅助网络关注面部结构。然而,卷积神经网络存在感受野大小受限的问题。此外,获取额外的人脸先验是一项艰巨的任务,并且不准确的人脸先验会严重影响生成图像质量。鉴于最近Transformer在视觉领域的出色表现,论文针对上述问题,提出一种基于Transformer与注意力聚合的人脸超分辨率重建网络来恢复人脸的全局结构与局部纹理细节。Most of the current face super-resolution(FSR)methods put the whole face image or additional facial prior into the convolutional neural network(CNN)to assist the network to focus on the facial structure.However,CNN suffers from the problem of restricted field size,which may destroy the consistency of the facial structure.In addition,obtaining additional facial priors is a tough task,and inaccurate facial prior can seriously degrade the quality of the generated images.Recently,in view of the outstand-ing performance of Transformer in the field of vision,this paper proposes a face super-resolution network based on Transformer and attention integration to recover the global structure and local texture details of the face.

关 键 词：人脸超分辨率 TRANSFORMER 注意力机制 残差块 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]