Models And Algorithms For Hyperspectral Image Restoration

Hyperspectral image(HSI)data are acquired by high spectral resolution sensors,and consist of hundreds of contiguous narrow spectral band images.With the wealth of available spectral information,hyperspectral image has been found to be very useful for many remote sensing applications,such as vegetation mapping,mineral exploration,urban planning and environmental monitoring.However,hyperspectral imaging sen-sors unavoidably introduce noise into the acquired HSI data during the imaging process,which severely degrades the quality of the imagery and limits the precision of the sub-sequent image interpretation processes,including unmixing,classification and target detection.Therefore,it is critical to reduce the noise in the hyperspectral image and improve its quality before the subsequent image interpretation processes.At present,the existing hyperspectral image denoising methods are widely using l1 norm to simu-late the non-gaussian noise including impulse noise、deadlines and stripes.The l1 norm can eliminate some sparse noise effectively,but still can produce biased estimation.So as to overcome this drawback,we study the model and algorithm by improving the sparsity.In order to remove the high level impulse noise,we propose an adaptive correction procedure for LRTDTV model,where the correction term is a linear term constructed from the initial estimate and aiming to improve the sparsity of norm.The process of adaptive correction is as follows:the first step is to generate the initial value by solving the LRTDTV model;The second step is the correction step,which may be repeated many times.At last,five kinds of experiment on simulated noise HSI data are carried out to demonstrate the superiority of the proposed method,especially for the removal of high level pepper and salt noise.Furthermore,we focus on the hyperspectral image denoising problem under the assumption that the signal is mainly corrupted by Guassian noise,impulse noise and strips.In order to fully utilize the structure of noiseless hyperspectral image,we propose LRMCPTV model based on tensor representation.In our model,the nonconvex and nonsmooth MCP function is adopted to remove sparse noise,the Tucker decomposition as a low rank constraint to further characterize the spectral similarity of all the pixels,and combine with the SSTV regularizer to explore the piecewise smooth structure in both spatial and spectral domains.At the same time,we adopt the 3-block ADMM algorithm to solve the model.Both simulated noise HSI data and real-world noise HSI data are carried out to demonstrate the superiority of the proposed method,especially for the removal of mixed noise.