基于实例归一化的图像超分辨率方法  

Image Super Resolution Based on Instance Normalization

作　　者：张鲁 李坤伦[1] 魏泽发 柴艳娜[1] ZHANG Lu;LI Kun-lun;WEI Ze-fa;CHAI Yan-na(Educational Technology and Network Center,Chang’an University,Xi’an Shaanxi 710064)

机构地区：[1]长安大学教育技术与网络中心,陕西西安710064

出　　处：《数字技术与应用》2021年第3期96-98,共3页Digital Technology & Application

摘　　要：卷积神经网络目前是图像超分辨率问题的主流研究方法,但常用的MSE损失函数重建图像缺乏纹理和细节,为了改善这一问题,本文尝试将实例归一化用于图像超分辨率,该方法依赖于训练过程中的每个样本而非整个批次,且需要配合使用大尺寸数据,实验结果表明,提出的方法在精度和视觉方面均优于VDSR等方法。Convolutional neural network is the mainstream research method of image super resolution.However,the reconstructed image by MSE lacks texture and detail.To solve the problem,this article attempts to use instance normalization in image super resolution,which relies on each sample in the training process instead of the entire batch,and requires large size training data.Experimental results show that the proposed method is better than VDSR and other methods in terms of accuracy and visual.

关 键 词：图像超分辨率 卷积神经网络 实例归一化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]