PROJECT TITLE :

Nonconvex-Sparsity and Nonlocal-Smoothness-Based Blind Hyperspectral Unmixing

ABSTRACT:

When it comes to hyperspectral data, blind unmixing (HU) is an essential technique for decomposing mixed images into constituent elements that are weighted by their respective fractional abundances. The use of nonnegative matrix factorization (NMF)-based algorithms has grown in popularity in recent years for this purpose and has shown encouraging results in testing. When it comes to abundances, a number of methodologies have been used to investigate the importance of two properties: sparseness and smoothness. As a result, all of the previous methods fail to take use of a natural hyperspectral image (HSI) quality known as non-local smoothness, which refers to the fact that identical patches in a broader HSI region all have the same smoothness structure. Previous attempts on other projects have shown that the use of this preceding structure can greatly increase the performance in this investigation of the HU problem. As the non-local total variation (NLTV) regularizer, we first look at such a prior in HSI. Furthermore, we generalise NLTV to non-local HSI TV (NLHTV) to make the model more suitable for the blind HU task by completely analysing the inherent structure of HSI. A non-convex log-sum form regularizer defining the sparseness of abundance maps is incorporated into the NMF model in order to suggest novel blind HU models (NLTV/NLHTV and log-sum regularised NMF, respectively) that incorporate these regularizers into the model. An efficient algorithm based on an alternative optimization strategy (AOS) and the alternating direction approach of multipliers is developed to solve the presented models (ADMM). For the blind HU challenge, the suggested solution outperforms other competing methods after extensive testing on both simulated and actual hyperspectral data sets.