Research On Hyperspectral Image Reconstruction And Super-Resolution Imaging Technique Based On Compressive Sensing

The spectral images can provide both spatial and spectral information of theobserved target, and are utilized for quantificational physical analysis andidentification on the observed target, thus they are widely applied in various fieldssuch as earth observation. Due to the huge demand for information, the resolution ofspectral images is becoming higher and higher. The resolution enhancement will leadto the nonlinear increases of manufacturing cost and system complexity. And a hugeamount of data are also induced which constitutes great pressure to computerprocessing, storage and transmission. Therefore it will conversely restrict thedevelopment and application of spectral images.Compressive sensing (CS) is a novel sampling theory. CS theory is proved thata small number of incoherent measurements of a compressible signal contain enoughinformation for exact reconstruction. Since the CS rate is far lower than the Nyquistrate, the requirements for sensors are greatly relaxed so that the problem of high costcaused by the blind pursuit of an excessively high-resolution sensor can be avoided.As one of the crucial issues, the CS reconstruction algorithm plays a key role in theapplication of CS theory and affects its practical implementation, hence it has been ahot study since it was presented. Under this background, this dissertation deeplystudied the CS reconstruction algorithms for hyperspectral images in order to findeffective and robust reconstruction algorithms, and also discussed theimplementation framework of compressive imaging. The main contributions andinnovation points of the dissertation are taken as follows:1. A CS reconstruction algorithm based on the side information is proposed forhyperspectral images in order to reduce the computational complexity. Sincehyperspectral images have strong interband correlations, a prediction algorithm isfirst applied to estimate the optimal prediction band of the current band fromprevious reconstructed neighboring bands. Then, the laplace model parameter of theprediction band and the current band is investigated. According to the variation ofthe laplace model parameter during the reconstruction iteration process, an improvedGPSR (gradient projection for sparse reconstruction) algorithm is proposed toreconstruct the current band. Since the prediction band can be regarded as the sideinfromation of the current band, the improved GPSR modifies the initialization and stopping criterion of the original GPSR. Experimental results show that the proposedalgorithm can provide both better reconstruction accuracy and higher computationalefficiency.2. A CS reconstruction algorithm based on the interband prediction and jointoptimization is proposed for hyperspectral images to improve the reconstructionperformance. First, we prove that the spectral correlations of compressivehyperspectral data are strong as that of the original hyperspectral images in theory.Then, a linear prediction algorithm is employed for removing the redundancy ofsuccessive hyperspectral measurement vectors. Last, the obtained residualmeasurement vectors with low entropy are recovered using the proposed jointoptimization based POCS (projections onto convex sets) algorithm with the steepestdescent method. In addition, a pixel-guided stopping criterion is introduced to stopthe iteration. Experimental results show that the proposed algorithm exhibits itssuperiority over state-of-the-art CS reconstruction algorithms at the samemeasurement rates, while with a faster convergence speed.3. A fast OMP (orthogonal matching pursuit) reconstruction algorithm isproposed for hyperspectral images. Because of the spectral structural similarity ofhyperspectral images, their sparse solutions in the transform domain have the samelocations of significant elements. On this basis, a CS reconstruction algorithm basedon joint sparsity model is proposed. In this algorithm, first, an adaptive spectral bandgrouping algorithm for compressive hyperspectral data is designed to divide themeasurement vectors into several groups and also to find out the optimal referenceband for each group based on least square criterion. Then, the reference bands arereconstructed using the OMP algorithm and the locations of significant elements foreach reference band are also recorded. Finally, all the bands except for the referenceband for each group are linearly reconstructed using the least square method usingthe recorded locations of corresponding reference band. Experimental results indicatethat the proposed algorithm can efficiently speed up the reconstruction process withreliable recovery quality.4. In order to solve the problems of long imaging time and high computationalcomplexity in compressive imaging, a novel framework for compressive imaging isproposed based on the zone control of DMD (digital micromirror device) and super-resolution (SR) reconstruction. At the end of sampling, a new measurement matrix isdesigned for zone control on DMD to enhance the acquisition accuracy of information. The new DMD with zone control is used to make compressivemeasurements. At the end of recovery, first, a low resolution image is recoveredfrom the received measurements by solving an optimization problem. Then, themodeling of an SR problem in a CS framework is constructed according to the zonecontrol process, and a total variation (TV) algorithm is exploited to solve such a CSbased SR problem for the original high resolution image. Experimental resultsdemonstrate that the proposed method can not only greatly shorten the imaging timebut also obtain excellent recovery performance with very low computationalcomplexity.