Research On Information Restoration Of Hyperspectral Image Based On Imaging Mechanism Analysis

The ultimate purpose of collecting hyperspectral image(HSI) is to interpret interesting information of ground from the area to which it corresponding. Essentially, such manner is indirect, since the information of ground that we want to acquire is firstly transferred into digital image by hyperspectral imaging device, and then analyzed and utilized by interpreter. Obviously, the fidelity of such information transformation process, i.e. the degree that the resulting image preserves the original information of ground plays a crucial role to the accuracy and precision of interpretation. However, in the imaging chain of HSI, there are some unavoidable factors, such as turbulence of atmosphere, vibration of platform of sensor, the imaging process always introduces deviations to the original information of ground, which reflects in the quality decreasing and information degradation of HSI. Restoring original information of ground from degraded HSI and increasing its interpretability is a key science problem in the fields of remote sensing application.In the entire imaging chain of HSI, information loss of input scene due to certain physical constraints of optical instruments and electronic devices of sensor is the major factor to the quality decreasing of HSI. To solve this problem, there are extensive documents proposed, and they achieved, at least to some extent, some achievements. However, most of them adopted some mathematical assumptions or approximations to imaging model, they only considered the image degradations as phenomenon, and neglected the physical nature underlying them, which is unreasonable. Actually, any types of degradation phenomenon of HSI can be ascribed to certain physical mechanisms in image formation process. Joint consideration such mechanisms and information restoration algorithms is beneficial to accomplish the conception innovations and theoretical breakthroughs of HSI information restoration, which has great guiding significance to the improvement of quality and interpretability of HSI. In this paper, we studied the physical mechanisms and restoration of information degradation occurs in sensor. In this dissertation, we majorly focus on sensor degradation model in imaging chain and its restoration, especially, we will study the unique features of degradation image when algorithm takes the actual sensor model into account, and how to eliminate such type of distortions. Concentrating on this topic, we will discuss following topics:Firstly, the information loss introduced by optical device of sensor and the corresponding restoration problem is studied. Most current deblurring algorithms of HSI assumed that degradation kernel leading to blurring as Gaussian function, however, it is the convolution of original scene and optical device(e.g. lens) that results in the blurring.To address this problem, we carried out some studies in this dissertation. Firstly, by theatrical analysis and experimental simulation, we validated that the hard cut-off manner of optical device will not only brings the blurring of spatial information of HSI, but also introduces ringing effect around edge area of image. Most of current works are unable to deal with such type of distortions. To solve such difficulty, we proposed a deconvolution algorithm based on Hessian-Schatten regularization. The proposed method represents two-order gradient information of image into Hessian matrix, extracts its eigenvalue and eigenvector,and then constrains the total energy via Schatten norm, which enables the regularization term to preserve high-order structure of image. Therefore, the proposed method is more beneficial to remove local distortion such as ringing effects.Secondly, in term of imaging system, the feature of downsampling of electronic device of line-scanning spectrometer and the corresponding restoration problem is studied.Most current super-resolution reconstruction algorithms implicitly assumed that the hyperspectral image they deal with is obtained by plane-scanning spectrometer, which is unreasonable due to the existence of line-scanning spectrometer. To solve this limitation,in this dissertation, we do some research on this topic. We firstly compared the imaging process of plane-scanning and line-scanning spectrometer, and analyzed the physical mechanisms underlying the processes, described the distinctions on the manner of imaging, and proved experimentally the inapplicability of current algorithms when they are used to restore line-scanning spectrometer; next, we studied the physical meaning of slitimage which is obtained by line-scanning spectrometer, and discovered that slit-image has characteristic of anisotropy; finally, aiming at the manner of imaging of line-scanning spectrometer and the anisotropy of slit-image, we proposed a novel framework of HSI super-resolution reconstruction, which restores HSI in slit-order. Moreover, to reduce the ill-posedness of solution of model under the new framework, we proposed a novel anisotropy regularization term, which further improves the performance of proposed algorithm.Thirdly, the restoration of spatial information distortion, such as jagged effects and broken lines, which is introduced by down-sampling process of electronic device of sensor, is studied. Most current algorithms involving this topic tried to remove those distortions in spatial domain, therefore they can’t catch the nature of such type of distortion. To oevercome such limitation, we analyzed the down-sampling process of electronic device in frequency domain, and discovered that such type of distortion is essentially the mixing of high-frequency components of image. Correspondingly, the scheme of reduction of such distortion should be how to remove those mixing components. To accomplish such goal, we expressed information restoration process as super-resolution problem in frequency domain, and proposed a fractal transforming based anti-aliasing super-resolution algorithm. The proposed method exploited the characteristic that the results of fractal transforming are independent of resolution of image, and the property that fractal coding has better performance when there are more self-similarity structures in the image.Since the proposed method captures the nature of information degradation, compared to traditional algorithm based on spatial domain, it has better effects in frequency domain.