| Compared with multi-spectral image,the biggest advantage of Hyperspectral image(HSI)is the particularly high spectral resolution,which depicts the nearly continuous spectral curve of target.It plays an important role in geological exploration,precise agricultural and forestry monitoring,national defense and military.But the acquired HSI often contain various types of noise,like Gaussian noise,impulsive noise,stripes or deadlines,etc.,due to the interference of various factors,like spectral instrument defects,atmospheric interference,calibration errors,etc.,during the acquisition.The types of noise in different bands are different,and its intensity is also different;many types of noise are mixed in the same band,which greatly limits the development of HSI’s subsequent applications,like target detection,recognition and classification,and brings challenges for noise reduction.This paper focuses on the difficulties of HSI denoising method.We have an in-depth study on HSI denoising methods based on low-rank representation and tensor decomposition.To design a HSI denoising method,we combined the low-rank structure of HSI with 3-D anisotropic total variation(3DATV)regularization.The main research contents of this article are as follows:(1)Many types of noise with different intensities exist in the same HSI.Some sparse noises are structural,such as stripes,that is,similar noise may exist in the same spatial position in different bands.And low-rank based denoising methods often regard structural noise as details,so it’s difficult to be removed;and some bands are seriously polluted by noise,which even covers many details,making it difficult to restore the structure features.To remove the mixed noise,enhance the performance of denoising structural noise and dense noise and improve the denoising effect,combined with the low-rank representation model,the Chapter 3 proposes a HSI denoising algorithm based low-rank representation and non-local similarity(NLRATV).The algorithm constructs 2D-Groups,increases the sparsity of sparse noise,weakens the correlation of structural noise,and reduces more structural noise.Meantime,it combines the local similarity and non-local similarity to increase structural constraint to protect more image structure.Then we use the 3DATV regularization to enhance the piecewise smoothness of the HSI and protect the image edge,texture and other details.(2)The NLRATV proposed in Chapter 3 is based on the matrix decomposition,and need transfer HSI into matrix,which splits its original structure.To protect the integrity of HSI,the Chapter 4 designs a HSI denoising algorithm based on non-local low rank representation and Tucker decomposition(NLRTDTV).The algorithm constructs 3D-Groups,uses non-local similarity to increase structural constraint,and extends NLRATV algorithm into the third-order tensor,using Tucker decomposition to denoising HSI integrally,so it protects the 3-D structure of HSI;3D-Group not only enhances the low-rank structure of the tensor,but also increases the sparsity of the sparse noise,so it enhances the ability of restoring band with serious noise and weakening of the structural sparse noise.In this paper,the effectiveness of NLRATV and NLRTDTV is verified by simulated and real data experiments.Experiments show that while denoised with NLRATV,the average peak signal-to-noise ratio,the average structural similarity,the energy gradient function and the variance are increased by 1.4%,1.4%,4.3%,and 3.2%.Basis on the results,NLRTDTV have increased by about 1.5%,1.1%,6.7%,and 3.1%,respectively.The performance of two algorithms on the average spectral angular distance index is relatively good.Simulated experiments show that the spectral curve of the image after denoising has a high coincidence with the original image,and the time cost of the algorithm is at the middle level of similar algorithms.Subjective and objective results show that the two algorithms improve the image quality,restore more details,and protect the structure features of image. |
PDF Full Text Request