Research On Denoising Algorithm Of Hyperspectral Image Based On Low Rank Theory

Hyperspectral imaging has developed rapidly in recent years.By joining imaging technology and spectral technology,the resulting images contain spatial and spectral information.It has been widely studied and applied in the fields of food safety,observation and survey of the environment,military security and satellite remote sensing.However,hyperspectral images are inevitably subject to various kinds of interference during imaging,transmission and storage.The interferences will bring various noises to the image,including Gaussian noise,impulse noise,dead line and stripes,which has a serious impact on the subsequent processing and data analysis.Therefore,denoising is an indispensable step.The denoising of hyperspectral images is usually taken as a preprocessing step before other hyperspectral image processing and analysis operations.Better classification and segmentation results can indeed be obtained from the denoised hyperspectral image.In order to achieve noise reduction,we usually need to depend on some prior information.The low-rank theory is a commonly used tool in recent years.Correlation algorithms often assume that the clean noise-free hyperspectral image is low-rank,and the noise part is sparse.Then we use mathematical methods to separate the low-rank part,which is the clean image,to achieve noise reduction.The existing noise reduction methods based on low-rank characteristics improve the noise reduction effect by introducing spatial information,but because they only use local similarity or non-local self-similarity,the sparse noise removal effect with certain structural information in the spectral dimension is poor.At the same time,In real hyperspectral images,there is usually not only one noise disturbance,but also a mixture of noises with different intensity and distribution in different bands.This allows some noise reduction methods to perform well in some bands of hyperspectral images but not so well in others.Therefore,based on the existing low-rank theory and their noise reduction methods,we have made some improvements in view of the shortcomings: the super-pixel block clustering algorithm comprehensively considers the local and non-local similarity of spatial information,and combines with the low-rank theory to achieve noise reduction;By establishing a band-by-band noise model,the mixed noise can be accurately estimated,and it is applied to the low-rank noise reduction model.Specifically,the research contents are as follows:1.A new hyperspectral image denoising method based on super-pixel block clustering and low rank characteristics(SCLR)is proposed.It joins local similarity and non-local self-similarity,which realizes the adaptive partition and clustering of blocks and avoids the dependence on strong prior information.The local details are well preserved while the non-local spatial self-similarity is fully utilized,and the homogeneous blocks composed of the clustered super-pixel blocks not only have good spatial low-rank attributes,but also have better spectral low-rank attributes.2.According to the different intensity and distribution of noise in different bands,a new band by band noise model and low rank characteristics(BNLR)noise reduction method is designed to integrate sparse noise and Gaussian noise.At the same time it achieve more accurate estimation of mixed noise.The removal effect of mixed noise is better improved.In order to verify the effectiveness of our methods in hyperspectral images,a certain amount of experiments have been carried out on Washington DC Mall and Indian Pines.And we compared other methods based on low-rank characteristics.The results show that our methods have better denoising performance in most noisy environments.