基于模糊核估计的图像盲超分辨率神经网络  

Blurred Image Blind Super-resolution Network via Kernel Estimation

作　　者：李公平 陆耀[1] 王子建 吴紫薇 汪顺舟 LI Gong-Ping;LU Yao;WANG Zi-Jian;WU Zi-Wei;WANG Shun-Zhou(School of Computer Science and Technology,Beijing Institute of Technology,Beijing 100081)

机构地区：[1]北京理工大学计算机学院,北京100081

出　　处：《自动化学报》2023年第10期2109-2121,共13页Acta Automatica Sinica

基　　金：国家自然科学基金(61273273);国家重点研究发展计划(2017YFC 0112001);中央电视台基金(JG2018-0247)资助。

摘　　要：模糊图像的超分辨率重建具有挑战性并且有重要的实用价值.为此,提出一种基于模糊核估计的图像盲超分辨率神经网络(Blurred image blind super-resolution network via kernel estimation,BESRNet).该网络主要包括两个部分:模糊核估计网络(Blur kernel estimation network,BKENet)和模糊核自适应的图像重建网络(Kernel adaptive superresolution network,SRNet).给定任意低分辨率图像(Low-resolution image,LR),首先利用模糊核估计子网络从输入图像估计出实际的模糊核,然后根据估计到的模糊核,利用模糊核自适应的图像重建子网络完成输入图像的超分辨率重建.与其他图像盲超分辨率方法不同,所提出的模糊核估计网络能够显式地从输入低分辨率图像中估计出完整的模糊核,然后模糊核自适应的图像重建网络根据估计到的模糊核,动态地调整网络各层的图像特征,从而适应不同输入图像的模糊.在多个基准数据集上进行了有效性实验,定性和定量的结果都表明该网络优于同类的图像盲超分辨率神经网络.Blind blurred image super-resolution is challenging and has important application values.This paper proposes a blurred image blind super-resolution network via kernel estimation(BESRNet),which mainly includes two parts:Blur kernel estimation network(BKENet)and kernel adaptive super-resolution network(SRNet).Given a low-resolution image(LR),the network uses the blur kernel estimation subnetwork to estimate the blur kernel from the input image,and then it uses the kernel adaptive super-resolution subnetwork to super-resolve the input lowresolution image.Different from other blind super-resolution methods,the proposed blur kernel estimation subnetwork gives the whole blur kernel,then the kernel adaptive super-resolution subnetwork dynamically adjusts the image features of different network layers according to the estimated blur kernel to adapt to different image degradations.In this paper,extensive experiments are carried out on multiple benchmark datasets.The qualitative and quantitative results show that proposed method is superior to other blind super-resolution methods.

关 键 词：模糊图像 模糊核估计 卷积神经网络 盲超分辨率 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]