Hyperspectral Image Mixed Noise Denoising With Multi-Attribute Joint Constraint

Hyperspectral image?HSI?has the advantages of high spectral resolution and rich texture information because of its with hundreds or thousands of spectral channels.The combination of spacial image and spectral bands provides favorable conditions for deep mining the ground surface features,and is widely used in resource exploration,urban planning,military monitoring and other important fields.However,during the process of capturing and transmitting,the HSIs are inevitably contaminated with various noises due to equipment limitation and atmospheric environment impact,which severely restricts the subsequent analysis and application.Hence,the task of eliminating the noise in HSIs is a valuable research topic in the field of remote sensing image processing.The denoising of the HSI is the process of recovering the clean image from the observed corrupted image,which is based on the noisy image and the data properties of HSI itself.Using the prior information of HSI to establish reasonable regularization terms is one of the effective manners to estimate the clean image.In this paper,a great number of prior information of the HSI is sufficiently exploited,such as local?global?low-rank,piecewise smoothness,and spatial sparsity.At the same time,the study on the removal of mixed noise in HSI is carried out,and relevant denoising algorithms are further designed.The main innovations of this paper are as follows:1.With the inspiration of the work L₀ gradient projection,theL₀ gradient constraint model is extended to measure the global sparsity of HSI.We present a novel HSI denoising approach by local low-rank matrix recovery withL₀ gradient constraint.The basic idea of this method is to effectively extract the essential signatures of HSI using a local low rank structure,and uses theL₀ gradient to constraint the sparseness of data to achieve the purpose of removing the mixed noise of HSI.In order to overcome the NP-hard problem caused by solving the non-convex discrete constraint,we derive a closed-form solution of the L₀ norm constraint and adopt the augmented Lagrange multiplier?ALM?method to calculate the approximate result of the nonconvex optimization problem.Extensive experiments on simulated and real HSIs demonstrate that the proposed method can maintain the sharp edge structure well and achieve highly competent objective performance.2.To compensate the shortcomings of residual noise based on the low rank method,we present a restoration method for HSI mixture noise removal via local low-rank matrix recovery and Moreau-enhanced total variation.Following the well-known total variation?TV?model for 2D image,we develop a Moreau-enhanced2D TV restoration model,which involves a non-convex penalty to maintain the convexity of the objective function.By incorporating the band-by-band Moreau-enhanced 2D TV denoising model into the local low-rank matrix recovery,a complete HSI mixed noise restoration framework is formed.The effectiveness of the proposed method is verified by extensive experiments.Experimental results indicate that this method can effectively suppress the mixed noise in HSI,preserve satisfactory spatial structure information,and effectively deal the troublesome problem of residual noise in the low-rank based model.3.For the purpose of reducing the computational burden,a Moreau-enhanced total variation combined with the subspace representation framework is proposed to recover noise-free HSI from mixed noisy counterparts.Based on the global spectral low-rank priors of HSI,we project the original image into a small size spectral dimension,and perform the Moreau-enhanced TV on each abundance map to separate the noise out.It is expected to achieve more performance boost of signatures through subspace factorization.Then,the noise-free HSI is obtained via the corresponding inverse transformation.Extensive results under simulated and real hypspectral images demonstrate that our method can not only protect the image details and reduce the spectral distortion,but also has a huge advantage in processing speed over the original spectral dimension data.