| With the wide application of hyperspectral imaging technology in biomedical imaging,remote sensing and astronomy,the algorithm of hyperspectral image reconstruction based on compressed sensing has become a hot research topic.Coded aperture snapshot spectral imaging(CASSI)recovers the 3D hyperspectral image from the corresponding 2D measurement.However,due to the undersamping which brings the serious data paucity problem,the accuracy of recovered hyperspectral image is usually limited.In this thesis,we propose a multitask nonparametric Bayesian dictionary learning algorithm for compressive hyperspectral image reconstruction under the blind compressive sensing,which is suitable for ab initio reconstruction problems with unavailable training datasets.Defining the compressive reconstruction of a 3D hyperspectral image from a 2D measurement captured by CASSI camera is a single task.Since multiple similar tasks are not statistically independent,taking advantage of statistical correlations among them can improve the performance of the inversion algorithm when these are performed simultaneously.The proposed algorithm jointly infers representation atoms and corresponding coefficients from highly undersampled measurements.The communication among multiple tasks is achieved by sharing a common dictionary and hyper-priors,which significantly alleviates the problem of data paucity.Utilizing the fully data-driven attribute of nonparametric Bayesian estimation strategy,the algorithm fulfills dictionary learning as well as image reconstruction from measurements rather than alien database.In addition,we consider the image-level sparse regularization,total variation(TV),which can alleviate the noise during the process of reconstruction to improve the accuracy of the reconstructed hyperspectral image and preserve the edge information of the image.In this thesis,we minimize TV via alternating direction method of multipliers(ADMM).What’s more,due to the inherent temporal redundancy in video sequences,the proposed algorithm can be extended to the temporal domain for reconstruction of compressed hyperspectral video.Based on the proposed method,an image reconstruction system is designed and developed for compressed hyperspectral image reconstruction.Experimental results demonstrate that the proposed algorithm has improved the reconstruction accuracy of hyperspectral images over other methods,and the TV penalty term enables reduce the noise. |
