抗噪声的分步式图像超分辨率重构算法  

Noise-resistant multistep image super resolution network

作　　者：郭昕刚[1] 何颖晨 程超[1] GUO Xin-gang;HE Ying-chen;CHENG Chao(School of Science and Engineering,Changchun University of Technology,Changchun 130012,China)

机构地区：[1]长春工业大学计算机科学与工程学院,长春130012

出　　处：《吉林大学学报（工学版）》2024年第7期2063-2071,共9页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金联合基金重点项目(U20A20186);长春市科技局重大专项项目(21GD05);吉林省科技厅重点攻关项目(20210201113GX)。

摘　　要：基于卷积神经网络的图像超分辨率重构方法大多数假设低分辨率图像是从高分辨率图像双三次降采样得到,而现实环境低分辨率图像带有未知噪声,不可避免地导致网络性能较差。针对这一普遍问题,本文提出了一种抗噪声的分步式图像超分辨率重构算法。首先,将信息蒸馏图像降噪网络结合生成对抗网络进行网络训练,以提高降噪网络的图像降噪能力;其次,将降噪网络的中间网络纯净特征图和降噪后的图像与分步式图像超分辨率重构网络结合,配合分步式网络训练,实现网络对真实环境低分辨率图像的有效超分辨率重构。在自建含有高斯噪声的BSD100^(*)与BSD100^(#)数据集上对本文提出的网络进行了训练和评估。实验结果表明:所提网络与已有先进网络相比,在图像质量评估和视觉对比上均取得较大提升。The image super-resolution reconstruction methods based on convolutional neural network mostly assume that a low resolution(LR)image is obtained by Bicubic downsampling of high resolution(HR)image.However,LR images in real environment contain unknown noise,which inevitably leads to poor network performance.To solve this common problem,A noise-resistant multistep image super resolution network is proposed.First of all,combine information distilling image denoising network and generative adversarial network to train the denoising network,in order to improve image denoising ability of the network;Secondly,the pure feature map in the middle layer of the network and the denoised image are combined with the stepwise image super-resolution reconstruction network,which is combined with the stepwise network training,to reconstruct low-resolution image of the real environment effectively.The proposed network is trained and evaluated on BSD100^(*)and BSD100^(#)datasets with gaussian noise.Experimental results show that the proposed network achieves good improvement in image quality evaluation and visual comparison compared with existing advanced networks.

关 键 词：深度学习 图像降噪网络 图像超分辨率重构网络 生成对抗网络 分步式网络训练 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]