摘要
现有的基于深度学习的图像盲超分算法仅利用神经网络端到端地学习低分辨率图像到高分辨率图像的映射,让网络隐式地学习图像的先验,导致算法仍产生模糊的超分结果。针对上述问题,提出一种稀疏性和自相似性先验引导的深度学习图像盲超分算法。首先,针对不同的低分辨率图像输入,利用动态线性核估计模块,有效估计出相应模糊核;然后,利用基于快速迭代软阈值收缩算法(FISTA)的深度展开反卷积滤波模块,显式地对信号的稀疏性先验进行建模,实现对退化图像的反卷积恢复;最后,利用双通道多尺度大感受野恢复模块,借助图像自相似性先验进行超分恢复。实验结果表明,相较于现有方法,所提出算法在公开的Gaussian8数据集上达到了31.66的峰值信噪比(PSNR)与0.8725的结构相似度(SSIM),在公开的DIV2KRK数据集上实现了29.08的PSNR与0.8007的SSIM,其所恢复出的图像不仅具有最高的复原指标,还具有更佳的视觉效果。
The existing deep learning-based image blind super-resolution algorithms only utilize neural networks to learn the end-to-end mapping from low-resolution(LR)images to high-resolution(HR)images,only allowing the net‐work to implicitly learn image priors,resulting in algorithms that still produce blurry super-resolution results.To ad‐dress the above issues,a deep learning image blind super-resolution algorithm guided by sparsity and self-similarity pri‐ors is proposed.Initially,for various LR image inputs,a dynamic linear kernel estimation module is employed to effec‐tively estimate the corresponding blur kernels;Subsequently,a deep unfolding deconvolution filtering module based on the Fast Iterative Shrinkage-Thresholding Algorithm(FISTA)is utilized to explicitly model the sparsity prior of signal,achieving deconvolution restoration of the degraded images;Finally,a dual-path multi-scale large receptive field resto‐ration module leverages the self-similarity prior of images for super-resolution recovery.The experimental results indi‐cate that,compared to existing methods,the proposed algorithm achieves a peak signal-to-noise ratio(PSNR)of 31.66 and a structural similarity index(SSIM)of 0.8725 on the publicly available Gaussian8 dataset,and attains a PSNR of 29.08 and an SSIM of 0.8007 on the DIV2KRK dataset.The images restored by the proposed algorithm not only exhib‐it the highest restoration metrics but also superior visual quality.
作者
葛孙逸
罗小伟
冯世阳
王斌
GE Sun-Yi;LUO Xiao-Wei;FENG Shi-Yang;WANG Bin(Key Laboratory for Information Science of Electromagnetic Waves(MoE),Fudan University,Shanghai 200433,China;Image and Intelligence Laboratory,School of Information Science and Technology,Fudan University,Shanghai 200433,China;Media Technology Resources Department,UNISOC(Shanghai)Technologies Co.Ltd,Shanghai 200120,China)
出处
《红外与毫米波学报》
北大核心
2025年第3期431-444,共14页
Journal of Infrared and Millimeter Waves
基金
国家自然科学基金(62371140)。
关键词
图像盲超分
深度学习
稀疏性先验
自相似性先验
深度展开网络
blind image super-resolution
deep learning
sparsity prior
self-similarity prior
deep unfolding network
