| Hyperspectral images(HSI)combines image information and spectral information of samples,which plays a huge role in remote sensing,medicine,agriculture,food and other fields.But the actual hyperspectral images often contain different types of noise such as Gaussian noise,pulse noise,fringe and dead line.The image information and spectral information destroyed by noise can not be used normally.Therefore,hyperspectral image denoising is a very important process.Traditional denoising methods,such as robust principal component analysis model(RPCA),approximate the rank function by introducing kernel norm,but this leads to the model is only sensitive to large singular values.The method based on smooth rank approximation not only treats each singular value equally,but also approximates the rank function more closely than the method based on kernel norm.Based on the idea of smooth rank approximation,this paper proposes the following three new denoising methods for mixed noise removal in hyperspectral images:(1)Hyperspectral image denoising via smoothing rank approximation for fusion dictionaries and spatial-spectral total variation(DSRASSTV): The traditional hyperspectral image denoising model lacks adaptability to various noises and effective utilization of spatial-spectral information.In order to solve these problems,the SSTV regularization and a new dictionary selection strategy is introduced into the smooth rank approximation model,so that the model has a better effect in removing mixed noises and recovering ground object information.(2)Hyperspectral image denoising based on group sparse and constrained smoothing rank approximation(CSRAGS): The constrained smoothing rank approximation(CSRA)is used to solve the problem caused by the nuclear norm,which can more accurately explore the internal low-rank structure of HSI.On the other hand,considering the empty spectral smoothness and internal group sparsity of HSI,a new group sparse regularization is added to the CSRA model.Group sparse regularization can make good use of the correlation between spatial and spectral dimension,improve the removal effect of mixed noise and ensure the smoothness of empty spectral dimension.(3)Hyperspectral image denoising based on constrained smoothing rank approximation and weighted E3DTV(SAWTV & SAWGS): The use of SSTV regularization is based on the prior that each band of the difference image has the same sparse distribution,while the real noisy HSI may not all meet the condition.To solve this problem,combined with the study of CSRAGS,the weighted E3 DTV regularization was introduced on the basis of the CSRA model,and two new denoising models were established by considering the sparsity and group sparsity within the image respectively.These models not only made the combination of low rank regularization and TV regularization more adaptive,but also had a better effect on the removal of complex noise interference. |
