Research On Sparsity Based Hyperspectral Images Processing Methods

Spectral imagery technologies have been greatly advanced due to the developments of the sophisticated sensors. Hyperspectral remote sensing, which can obtain fine-grained spectral information in addition to acquiring spatial information of the images, had increasingly gained popularity in the fields such as military surveillance and economy development. However, with the improvement of the hyperspectral image resolution, the data obtained by spectral imager had greatly been beyond the capacity of transferring and processing of the devices. Sparseness constraints based signal processing methods have been the research focus recently, and it have been widely applied in signal processing, pattern recognition and computer vision, etc.Aiming at difficulty of processing hyperspectral imagery, this dissertation analyses the sparsity of the hyperspectral images and study the hyperspectral image classification and object recognition based on the sparseness constraints. The works of the dissertation is as follows:First, the dissertation analyses and verifies the sparsity of the hyperspectral images. Analyzing the typical characteristics of the hyperspectral image data, the dissertation demonstrate the sparsity of the hyperspectral images by undergoing sparse decomposition on hyperspectral data with unsupervised dictionary learning and then comparing the real spectral curves with ones based on the unsupervised dictionary learning, which shows the two curves can match to each other very well.Second, the dissertation proposes a sparseness embedding based hyperspectral image classification method. Considering the character of high dimensionality of hyperspectral images, the method extract the features of hyperspectral images by utilizing sparseness embedding and undergo within-class sparse reconstruction on the condition of preserving the within-class compactness and maximizing the between-class scatter to enhance discriminative ability of the features in the projected space. The experiments based on the real data show that the proposed method can classify the data of multiple classes and improve both on the classification time consuming and the classification accuracy comparing with other methods.Finally, the dissertation further analysis the anomaly detection problem of the hyperspectral images and propose the accumulating pyramid spatial-spectral collaborative coding divergence for hyperspectral anomaly detection. A simple and efficient unsupervised sparse feature learning method, which works by optimizing the population sparsity, lifetime sparsity, and high dispersal of feature distribution, is firstly adopted to extract spectral features from the original high dimensional spectral vectors. Spatial pyramid local window selection strategy is then introduced to encapsulate feature variations of local neighborhoods with different sizes. Finally, the spatial-spectral collaborative coding divergence over multiple levels is fused to localize potential anomalies. In the end, this chapter conducts experiments about the proposed method. Experimental results demonstrate the proposed method is robust and efficient for real-time applications.