Research On Information Restoration Method Of Hyperspectral Images

With the continuous development of remote sensing and imaging spectrometry technology,not only the limitation of observation distance is broken,but also the spectral constraint of visible image is get rid of.Hyperspectral images(HSI),with higher spectral resolution,build an information bridge between the composition of microcosmic particles and the distribution characteristics of macroscopic objects.The image-spectrum merging capability makes remote sensing break through the bondage of vision,closer to physical reality and widely use in the fields of resource exploration,agricultural detection,urban planning,target identification.However,the hyperspectral imaging chain is complex,the degradation introduced by factors such as sensors and atmosphere will affect the spatial-spectral characteristics and limit the perception of the ground objects.At present,limited by physical principle and hardware,the degradation can not be fundamentally avoided in a short period of time;meanwhile,the existing platforms produce and accumulate massive HSI data,and quality reduced data still exist and have great research value.Therefore,the restoration of damaged HSIs based on digital image processing technology is of great significance for guiding industrial and agricultural production facilities and national defense construction.Due to the image-spectrum merging characteristics,any degradation will cause spatial and spectral dimension damage simultaneously.On the one hand,traditional methods usually ignore the spatial-spectral characteristic protection when restoring damaged information;on the other hand,the causes of various degradation are complicated,presenting different statistical property and are difficult to restore effectivly by a single method.The core task need to be solved ugently is to realize the systematic restoration of massive HSIs with unknown quality,and make the data more accurately reflect the spatial-spectral radiation information.Based on the principle of electromagnetic radiation and the basic mechanism of spectral imaging,this paper breaks the limitation of digital image processing technology frame,and proposes a new thoughts of comprehensive recovering multiple degradation by utilizing spatial-spectral correlation of HSIs,prior statistical characteristics and multi-source remote sensing information.Firstly,the imaging mechanism of HSIs is highly related to characteristics of the images and the degradation,and the basis of designing the restoration method distinguish to conventional image processing technology.This paper starts with the basic principles and characteristics of electromagnetic radiation,explains the imaging mechanism of hyperspectral remote sensing from wave property of light,analyzes the correlation between spectral characteristics and the reflection/radiation characteristics of ground objects from particle property of light.On this basis,continuous bands contaminated by water absorption and particle scattering in the atmosphere(junk bands)are further analyzed,and the corresponding degradation models and detection methods are introduced.The derivative of the spectral curve is discussed from microcosmic and macroscopical perspectives,and a restoration method based on bidirectional derivative spectrum prediction is proposed.The superiority of the method is proved from a mathematical viewpoint using Taylor series expansion,and experiments show that the method can better maintain the diagnostic characteristics of the image.Secondly,an edge protected adaptive morphological filtering algorithm is proposed aiming at restoring stripped dead pixels caused by spatial dimension scanning in imaging.Utilizing the spatial-spectral correlation,we first extract edges by calculating spatial gradient from spectral dimension to describe the structure information and detect the damaged area.Then,edges in the damaged area are reconstructed based on edge neighborhood probabilistic support and learning algorithm to estimate the missing spatial structure information.Finally,structuring elements restrained by edges are generated,combining with adaptive morphological filtering,the proposed method can recover the spatial and spectral information in stripes.Experiments show that our method can better maintain the spatial structure characteristics and effectively restrain the spectral aliasing at regions boundaries.Besides,through a decision level postprocessing experiment,the proposed adaptive morphological structuring elements can provide a universal module to improve the spatial characteristics protection ability of various filtering algorithms,which is of great value for the processing and intelligent interpretation of HSIs.Thirdly,we propose a restoration method based on super pixel segmentation and multiresolution low rank expression aiming at eliminating mixture noises.In the process of photoelectric conversion,signal processing and transmission in sensors,there always exsits mixture noises such as Gaussian noise,sparse noise and structural noise and so on.We first analyze the statistical characteristics of the clean data and noisy component,establish the degradation model.Since the global low rank property of HSIs is not fully used in existing methods,a multiresolution low rank restoration model is established in this paper.The structural noise with local low rank property is suppressed in the low resolution layer;in the high resolution layers,a linear spatial-spectral clustering algorithm is used to obtain the space adjacent and spectral similar super pixels instead of the traditional blocks as the low rank recovery units,both the stability of the species and the low rank property in each unit are improved.Experiments show that the proposed method can better suppress the mixture noises,protect the spatial-spectral characteristics of HSIs and effectively solve the block effect problem in the traditional low rank based restoration algorithms.Finally,focuses on the spatial and spectral resolution tradeoff caused by the bandwidth limitation of the current imaging spectrometer device information conversion,a local adaptive sparse unmixing and sub-pixel calibration based fusion superresolution algorithm is proposed.The resolution enhanced HSIs are obtained by using the spatial and spectral correlation between hyperspectral and multispectral images.According to the probability distribution of ground objects,the sparsity of the abundance matrix is introduced into the coupled spectral unmixing model,and the multiplicative update for nonnegative matrix factorization is optimized;then high spatial-spectral resolution HSIs with better characteristics are more effectively obtained,and the convergence of the algorithm is obviously improved.After that,in order to further improve the spatial accuracy of the fused images,an optimal matching based adaptive morphology filtering method is proposed aiming at subpixel calibration for biases introduced by multi-source data registration and fusion process,the spatial characteristics of the target fused data should be as close as possible to multispectral images.Experiments show that the acquired fusion superresolution HSIs are more accurate in representing the intrinsic spatial-spectral characteristics of ground objects than the existing advanced methods.