本文關鍵詞：基于核稀疏和空間約束的高光譜目標檢測方法研究　出處：《哈爾濱工業(yè)大學》2016年碩士論文　論文類型：學位論文

  更多相關文章： 高光譜目標檢測 稀疏表示 核方法 空間信息 聯(lián)合檢測

【摘要】：高光譜遙感是一種能對地物進行精細觀測的信息獲取手段,高光譜圖像的目標檢測具有較高的研究價值。在軍事應用方面,主要用于對重點目標進行檢測與識別;在民用方面,高光譜目標檢測在環(huán)境監(jiān)測、礦產(chǎn)資源定位、精細農(nóng)業(yè)等領域也有廣泛的應用。論文主要針對高光譜數(shù)據(jù)的典型特征,在信號稀疏表示理論的基礎上綜合利用其蘊含的光譜信息和空間信息,實現(xiàn)高光譜圖像的目標檢測。論文首先從信號稀疏表示的基礎理論起步,以信號稀疏表示的數(shù)學模型為前提,明確了使用范數(shù)松弛求解稀疏方程時數(shù)據(jù)應滿足的條件,并給出求解稀疏方程的基本方法。后續(xù)分析高光譜圖像具有的稀疏性、“圖譜合一”等特征,將稀疏表示模型應用于高光譜圖像數(shù)據(jù),由信號的稀疏重構過渡到基于稀疏表示的高光譜目標檢測方法,并給出基于實驗數(shù)據(jù)的高光譜目標檢測初步結果。其次,由于非線性和“圖譜合一”是高光譜圖像具有的重要特征,因此著重研究光譜與空間信息在高光譜目標檢測中的優(yōu)化利用方法。針對高光譜數(shù)據(jù)的非線性特征可能導致的線性目標檢測算法的欠佳結果,使用核方法對基于稀疏表示的高光譜目標檢測基本方法進行改進,使其具有處理非線性數(shù)據(jù)的能力,從而優(yōu)化利用高光譜數(shù)據(jù)中的光譜信息。針對高光譜數(shù)據(jù)蘊含的空間信息,使用邊緣描述和水平集描述對目標對象的空間特征進行提取,從而深入挖掘高光譜數(shù)據(jù)圖像域信息,進而為后續(xù)的空間信息應用提供數(shù)據(jù)基礎。最后,考慮高光譜數(shù)據(jù)中目標對象光譜信息和空間信息的統(tǒng)一性和整體性,建立光譜-空間信息聯(lián)合稀疏模型,力求更為全面準確地描述目標對象。在聯(lián)合稀疏模型的基礎上,借助空間域信息輔助目標檢測,提出了基于光譜信息的目標檢測結果的修正算法。論文利用AVIRIS高光譜數(shù)據(jù),研究了稀疏度和核參數(shù)對目標檢測結果的影響。同時,由正常數(shù)據(jù)集衍生構建了弱完整性數(shù)據(jù),并在正常數(shù)據(jù)集和弱完整性數(shù)據(jù)集上測試經(jīng)典目標檢測方法、基于稀疏表示的目標檢測方法和基于光譜-空間信息聯(lián)合稀疏模型的目標檢測方法,現(xiàn)有數(shù)據(jù)條件下的實驗結果表明,利用高光譜數(shù)據(jù)的光譜和空間信息的目標聯(lián)合檢測方法性能最為優(yōu)越,對弱完整性數(shù)據(jù)的適應性也更強。
[Abstract]:Hyperspectral remote sensing (hyperspectral remote sensing) is a kind of information acquisition means for precise observation of ground objects. The target detection of hyperspectral images has high research value. In military applications, it is mainly used for detecting and identifying key targets. In civilian aspect, hyperspectral target detection is widely applied in environmental monitoring, mineral resource location, precision agriculture and other fields. Aiming at the typical characteristics of hyperspectral data, based on the theory of sparse representation, the paper uses the spectral information and spatial information contained in hyperspectral data to achieve target detection in hyperspectral images. Starting from the basic theory of sparse representation, the paper first defines the mathematical model of signal sparse representation, and specifies the condition that data should satisfy when using norm relaxation to solve sparse equations, and gives the basic method to solve sparse equations. Subsequent analysis of hyperspectral image is sparse, "one map" and other characteristics of the sparse representation model is applied to hyperspectral image data, hyperspectral target detection method based on sparse representation by sparse reconstruction signal is given to the transition, the preliminary results of hyperspectral target detection based on experimental data. Secondly, due to the nonlinear and "spectral unification" is the important feature of hyperspectral image. Therefore, the optimal utilization of spectral and spatial information in hyperspectral target detection is mainly studied. The poor results of linear target detection algorithm for the nonlinear characteristics of hyperspectral data may lead to the use of nuclear methods to improve the basic method of hyperspectral target detection based on sparse representation, which has the ability to deal with nonlinear data, so as to optimize the use of spectral information in hyperspectral data. Aiming at the spatial information contained in hyperspectral data, edge description and level set description are used to extract the spatial characteristics of target objects, so as to dig out hyperspectral data and image domain information, and provide data basis for subsequent spatial information application. Finally, considering the uniformity and integrity of the spectral and spatial information of target objects in hyperspectral data, we establish a joint sparse model of spectral spatial information, and strive to describe target objects more comprehensively and accurately. On the basis of joint sparse model and space domain information aided target detection, a correction algorithm for target detection results based on spectral information is proposed. Using AVIRIS hyperspectral data, the influence of sparsity and kernel parameters on target detection results is studied in this paper. At the same time, the construction of weak integrity data derived from the normal data set, and the normal data set and weak integrity data set test method, the classical sparse representation of the object detection method and detection method based on spectral spatial information combined with sparse model based on existing data conditions. The experimental results show that the target joint the detection method uses the spectral and spatial information of hyperspectral data is the most superior performance, the weak data integrity adaptable.
【學位授予單位】：哈爾濱工業(yè)大學
【學位級別】：碩士
【學位授予年份】：2016
【分類號】：TP751
，

本文編號：1342738