Research On Hyperspectral Imagery Sparse Target Detection Algorithms

Hyperspectral remote sensing is a multi-dimensional information acquisition technology combining spectral analysis and image processing,and it can not only get spatial information but also widely covered spectrum information.This characteristic makes it possible to understand an image by both morphological information and spectral information which is more extensive.In hyperspectral imagery(HSI),spectral reflectance data of different objects are completely different,and this is the theoretical basis of HSI processing technology.It means that either detection,classify,or other processing of HSI can be based on spectral information.Once a complete spectral library is obtained,matching method which is a simple binary classification issue can be used to detect targets,as long as the matching result can point out the pixel is target or not.Sparse representation algorithm is a hot topic in recent years,and its features can just meet the requirement of HSI target detection.So in this paper,a combination and some improvements of these two technologies are proposed.The main work of this paper is summarized as follows:At first,since HSI contains extremely wealthy information,the amount of HSI data is huge.To solve this problem,this paper makes a combination of HSI processing technology and sparse representation algorithm on the basis of a detailed study of both.The combined algorithm can use the high sparsity of HSI,at the same time,it reduced the arithmetic pressure brought by the large amount of data of HSI.The result of the comparative experiments of traditional HSI target detection algorithm and the sparse HSI target detection algorithm shows that the sparse one is more effective.Secondly,improve the process of solving sparsity coefficient which is a key process of sparse representation.Common sparsity coefficient solving algorithms are generally not effective enough,and they often take a very long operation time when faced large data such as HSI data.This paper proposed using StOMP algorithm to solve sparsity coefficient,and dramatically improved the operation speed under the premise of no significant decrease of detection accuracy.Comparative experiments verify the effectiveness of the proposed algorithm.At last,this paper presented an accurate sparse HSI target detection algorithm without prior information through a new method for constructing the dictionary.Previous dictionaryconstruction methods can be divided into two categories: One category is choosing a certain number of pixels from text HSI to contribute dictionary.The other one is extending over-complete dictionary by using prior knowledge of the text HSI.Both categories have their own pros and cons: the first one do not need prior knowledge,but because of the limitation of the number of chosen pixels and the deficiency of choosing method,information contained in the dictionary is often inadequate,and the detection effect is not ideal.The latter one have a better detection effect than the former one.But in practical applications,it is very hard to get complete prior knowledge,so that this kind of methods' practicality is not strong enough.For these two issues,this paper proposed a over-complete dictionary construction method using the result of RX anomaly detection as prior information.This method cleverly combines the advantages of both traditional methods,and avoids their disadvantages.Comparing experimental results show that the unsupervised dictionary construction method proposed in this paper has a high detection accuracy.