基于深度先验和小波变换的遥感图像复原算法

Remote Sensing Image Restoration Algorithm Based on Deep Prior and Wavelet Transform

针对各种退化因素导致遥感图像模糊的问题,在混合即插即用(Hybrid Plug-and-Play,H-PNP)模型的基础上,设计了基于深度先验和小波变换的模糊遥感图像复原算法。首先,利用导向滤波对模糊图像进行预处理,再构建结合非局部相似块低秩先验、深度先验和小波变换的模糊遥感图像复原模型,最后利用交替迭代法求解模型,复原出清晰的图像。考虑到惩罚因子对图像复原的影响,在深度去噪器中引入相对残差自适应原则更新惩罚因子。实验结果表明,上述算法对于叠加模糊和噪声的退化图像具有良好效果,选取GSR算法、NSCR算法和H-PNP算法进行对比实验,在主观视觉效果、峰值信噪比、特征相似性和结构相似性方面均优于对比算法。

Aiming at the problem of blurred remote sensing images caused by various degradation factors,this paper uses the Hybrid Plug-and-Play algorithm to design a fuzzy remote sensing image restoration algorithm based on deep priors and wavelet transform.The algorithm first uses guided filtering to process the blurred image and then constructs a restoration model of blurry remote sensing image that combines the non-local similar block low-rank prior,deep prior and wavelet transform.An alternate iteration method is used to solve the model.Considering the influence of the penalty factor on image restoration,a relative residual adaptive principle is used in the deep denoiser to update the penalty factor.The experimental results show that the algorithm has a good effect on degraded images with superimposed blur and noise.This paper selects the GSR algorithm,NSCR algorithm and H-PNP algorithm to be comparative experiments.The results illustrate that the algorithm is superior to comparative algorithms in aspects of subjective visual effect,peak signal-to-noise ratio,feature similarity and structural similarity.