Search

Compressive hyperspectral imaging (CHI) has been attracting great interests recently. To reconstruct hyperspectral (HS) images precisely from very few measurements is the main issue considered in CHI field. In this paper we propose a novel and extensible CHI approach including sampling strategy and the corresponding reconstruction method. The spatial energy distribution information (SEDI) can be explored during sampling process where the additional sampling time and measurements created by SEDI are negligible. We develop an SEDI based reconstruction method aiming to eliminate abnormal spikes and correct energy distortion which usually exist in reconstructed HS images. Simulations with real HS data demonstrate the fact that the proposed approach performs better than that without SEDI.

In the sparse Photoacoustic Microscopy system, a uniform random sampling scheme and low-rank matrix approximation-GoDec algorithm have been proposed for fast data acquisition and image recovery. However, this low-rank approximation algorithm leads to low resolution and fuzzy structure details. In this paper, the weighted Singular Value Thresholding algorithm is first applied in sparse PAM system to recover PAM images with high resolution. An efficient iterative weighting scheme is used for exactly solving matrix completion problem. Therefore, this algorithm is more accurate than classic SVT algorithm, which is proved by singular value analysis. Moreover, both simulations and real data experiments verify the performance of the weighted SVT algorithm and the application potential of this modified sparse PAM system.

The noise removal and endmember unmixing are two pivotal problems for the hyperspectral image (HSI) data exploitation. Different from the traditional approach that treating these two schemes separately, the proposed coupled Denoising and Unmixing (cDeUn) model integrates them into a uniform framework. To obtain the desired denoising and unmixing performances, three issues are included. First, the weighted nuclear norm minimization (WNNM) is adopted to separate the clean HSI from the noisy observation. Second, the sparse representation (SR) based unmixing scheme is incorporated to obtain the endmember abundance matrix. Both denoising and unmixing schemes act as constraints to each other and the effects of them can promote each other. Third, the hypergraph regularization is incorporated to retain the smooth consistency of the abundance matrix and further improve the final performance. Both synthetic and real data experiments confirm that the proposed cDeUn approach can provide superior denoising and unmixing performances for the HSI corrupted by the mixture of Gaussian, salt and pepper, and deadline noise.

Human-like musculor-skeletal arm model with its unique smooth and natural movements as well as low energy cost has attracted many researchers' interests and developed rapidly nowadays. In this paper, we focus on a human-like musculor-skeletal arm model with three links driven by nine muscles and propose a control model with two neurocontrollers trained by the NEAT algorithm. The NEAT (Neuroevolution of Augmenting Topologies) is a novel neuroevolution algorithm using a modified genetic algorithm to train neural networks by changing the topology structure as well as connection weights simultaneously. To evolve the network, we generate the training sets by using forward kinematics, geometry relationships, and muscle mechanic equations. By using two neurocontrollers, Position-Angle and Angle-Activation, the control model can deal with the redundancy and non-linear problems in separate neurocontroller at the same time.

This paper addresses the image sequences reconstruction by l 0 minimization, which is an NP-hard problem with high computational complexity. To solve it, we propose a novel recursive intelligent matching pursuit(RIMP) algorithm in this paper. The idea of RIMP lies in utilizing the recursive reconstruction to reduce the sparsity and applying the superiorities of intelligent optimization algorithm to solve the l 0 minimization essentially, which is beneficial to improving the reconstruction accuracy. To reduce the computational complexity, some matching strategies of greedy algorithm are used to design the intelligent searching strategies to accelerate the reconstruction speed. Also, based on the high reconstruction accuracy of the intelligent searching, RIMP can significantly improve the reconstruction accumulation in recursive reconstruction for image sequences. Numerical simulations on several image sequences have been used to demonstrate that the theoretical reconstruction performance of RIMP can be achieved.

The texture feature is an important information for hyperspectral image classification. In this study, we extend the traditional 2D GLGCM(gray-level gradient cooccurrence matrix) into 3D GGGTF(gray-gradient-gradient tensor field), which can extract gray and gradient texture features of hyper-spectral volume data simultaneously. A few statistical features are extended into third-order forms in order to calculate texture properties of the generated GGGTF. And then, the extracted texture features are classified by linear polynomial kernel SVM classifier. Two widely used hyperspectral datasets are used to test the performance of the proposed GGGTF. Experimental results demonstrate that it outperforms traditional 2D GLGCM method in feature extraction for supervised classifications.

The sparse signal recovery problem has been the subject of extensive research in several different communities. Tractable recovery algorithm is a crucial and fundamental theme of compressive sensing (CS), which has drawn significant interests in the last few years. In this paper, we firstly analyze the iterative residual in Orthogonal Matching Pursuit (OMP) algorithm. Secondly, a greedier algorithm is introduced, which is called Greedy OMP (GOMP) algorithm. This algorithm iteratively identifies more than one atoms using greedy atom identification, and then discards some atoms, which are of high similarity with the optimal atom. Compared with OMP algorithm, the experiments conducted on Gaussian and Zero-one sparse signal demonstrate that the proposed GOMP algorithm can provide better recovery performance. Finally, we experimentally investigate the effect of greedy constant in GOMP upon the recovery performance.