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

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.