基于ViT-CNN特征增强的图像超分辨率  

Image super-resolution based on feature enhancement with ViT-CNN

作　　者：高志荣[1] 孙清清 熊承义[2,3] 李帆 郑瑞华[2,3] GAO Zhirong;SUN Qingqing;XIONG Chengyi;LI Fan;ZHENG Ruihua(South-Central Minzu University,College of Computer Science,Wuhan 430074,China;South-Central Minzu University,College of Electronic and Information Engineering,Wuhan 430074,China;South-Central Minzu University,Hubei Key Lab of Intelligent Wireless Communication,Wuhan 430074,China)

机构地区：[1]中南民族大学计算机科学学院,武汉430074 [2]中南民族大学电子信息工程学院,武汉430074 [3]中南民族大学智能无线通信湖北省重点实验室,武汉430074

出　　处：《中南民族大学学报(自然科学版)》2025年第2期253-259,共7页Journal of South-Central Minzu University(Natural Science Edition)

基　　金：多谱信息处理技术国家重点实验室基金资助项目(6142113210303);中央高校基本科研业务专项资金资助项目(CZY21013)。

摘　　要：卷积神经网络(CNN)可以提取图像的局部相关特征,视觉Transformer(ViT)则侧重于捕获图像的远距离依赖关系,二者有效结合能够改进图像的重构质量.研究了一种基于ViT-CNN特征增强的图像超分辨率(SR)网络.具体来说,网络包含了基于ViT的SR分支与基于CNN的梯度分支,SR分支主要用于提取图像特征域中的全局相关性,而梯度分支则专注于图像梯度域中的局部依赖关系.通过对两种信息的融合与渐进增强,获得高倍放大的重构图像.此外,在网络的学习阶段引入了梯度损失及渐进训练策略,有效降低了网络的训练难度并增强了训练的稳定性.在多个公开数据集上的大量实验结果验证了所提方法在改善重构系统性能方面的有效性.The effective combination of Convolution Neural Network(CNN)which extract the local correlation features of images and Vision Transformer(ViT)which focuses on capturing the remote dependence of images can improve the quality of image reconstruction.A network of image super-resolution based on feature enhancement with ViT-CNN is studied.Specifically,the network includes ViT-based SR branch and CNN-based gradient branch,which extract the global correlation in the image feature domain and the local dependency in the image gradient domain respectively.Through the fusion and gradual enhancement of the two kinds of information,the reconstructed image with large factor is obtained.In addition,by introducing gradient loss and progressive training strategy,the difficulty of training is effectively reduced and the stability of training is enhanced.A large number of experimental results on multiple public datasets demonstrate the effectiveness of the proposed method in improving the performance of the reconstruction system.

关 键 词：图像超分辨率 卷积神经网络 视觉Transformer 特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]