基于拉普拉斯金字塔网络的融合降噪模块图像超分辨率算法  

Laplacian pyramid network fusion denoising module for image super resolution

作　　者：郭昕刚[1] 何颖晨 程超[1] GUO Xingang;HE Yingchen;CHENG Chao(School of Computer Science&Engineering,Changchun University of Technology,Changchun 130102,China)

机构地区：[1]长春工业大学计算机科学与工程学院,吉林长春130102

出　　处：《长春工业大学学报》2023年第4期335-344,共10页Journal of Changchun University of Technology

基　　金：吉林省科技厅重点攻关项目(20210201113GX);长春市科技局重大专项(21GD05)。

摘　　要：针对大多数图像超分辨率网络对带有未知噪声的低分辨率图像进行超分辨率重构的效果较差,提出一种拉普拉斯金字塔网络融合降噪模块的图像超分辨率算法。以LapSRN为基础融合降噪模块,并采用递归模块,同时引入TVLoss与降噪损失的LFDSR网络,在保持网络轻量级的基础上,提高重构后图像的清晰度。实验结果表明,文中算法输出图像更逼真和清晰,在抑制输入图像中的噪声对网络输出图像的过度影响中取得较好效果。Most of the current image super-resolution(SR)network have poor effect on the reconstruction of low resolution images with unknown noise.Therefore,an image super resolution algorithm based on Laplacian pyramid network is proposed.An LFDSR network based on LapSRN fusion denoising module,and a recursive Module is adopted,is proposed to improve the resolution of SR reconstruction model for LR input images in unknown environments,while maintaining the lightness of SR reconstruction model.TVLoss and denoising loss are introduced.Experimental results shows that the proposed network is more realistic and clearer than the advanced SR reconstruction network,and achieves better results in suppressing the excessive influence of noise in the input image on output image of network.

关 键 词：图像超分辨率重构 图像降噪 LapSRN 残差块 递归网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]