本研究致力於發展一套平行運算方法，針對高光譜影像做異常物 偵測與相異物質區別；透過個人電腦叢集 (PC cluster) 環境運算，達到有效提升偵測速率的目的。 我們的偵測方法主要概念，如果將影像中的背景視為高斯分佈 (Gaussian distribution)，由於異常物 (anomalies) 或人造物的光譜特性和背景有著極大差異且面積相對背景來說較小，因此可藉由找尋和背景分佈的不同來達到異常物偵測目的。若資料是一個高斯分佈，代入偏度 (skewness) 和峰度 (kurtosis) 的式子後結果應為零，如果資料非高斯分佈，利用偏度與峰度便可偵測有無異常物的存在。 結合偏度和峰度運算與正交子空間投影 (Orthogonal Subspace Projection, OSP) 概念，可達到偵測多類異常物之目的；不過受限於運算流程，提供一個初始投影行向量在一個時間內只能找出一類異常物，欲找出其他異常物，則需等待前一個異常物被找出後才能再進行。然而給定不同的投影行向量將可偵測出不同的異常物，因此本研究提出平行化的運算方法，透過訊息傳遞介面 (Message-Passing Interface, MPI) 的函式庫在個人電腦叢集上進行平行運算，將一個初始投影行向量傳送給一個CPU，讓一個CPU找出一類異常物，使具有多個CPU的叢集電腦在同一時間內找出多類異常物，大大提高偵測效率。 實驗利用美國太空總署噴射推進實驗室 (NASA’s Jet Propulsion Laboratory) 所發展的AVIRIS (Airborne Visible / InfraRed Imaging Spectrometer) 高光譜影像做測試，評估此平行化異常物偵測與識別演算法的運算效能。在兩套分別擁有四個節點 (4 Nodes) 但硬體性能相異的個人電腦叢集環境上執行平行演算法，並觀察比較不同運算環境下的效能差異。相較於在單台個人電腦上的運算時間，四節點叢集電腦可達到平均約2.7倍的速度提升，且保有原循序作法相同的物質偵測結果。 The aim of this thesis is to develop a parallel version of the high-order statistic algorithm for anomalies detection and discrimination in hyperspectral images. This approach was implemented by C language with the message passing interface (MPI) library on PC cluster. Anomaly detection is to detect unknown small targets from an unknown background. As a result, anomaly detection in an unknown image scene can be accomplished by searching the deviation from background distribution. It is known that skewness and kurtosis are the normalized third and fourth central moments which can be used to measure the asymmetry and flatness of the data distribution respectively. Since Gaussian distribution has zero skewness and kurtosis, if the distribution of the image has high skewness or kurtosis, it cannot be modeled as Gaussian and there are some anomalies resident in the image scene. Therefore anomaly can be detected by finding the maximum skewness or kurtosis direction. Using skewness or kurtosis with orthogonal subspace projection, anomaly discrimination can be achieved. Limited by operation procedure, a projector can only detect one anomaly at a time and the next procedure has to wait until the previous one had finished. Hence we propose a parallel version of the high-order statistic algorithm to distribute the computation loading. In our approach, each node will process one randomly initialized projector, and it will detect one anomaly target when the projector had converged. By means of this measure, the PC cluster is able to search for more than one anomaly at the same time. The algorithm then collects and analyzes the projectors from all nodes, and redistributes them to all nodes to search for the next possible anomalies. The computational performance of the proposed parallel anomaly detection algorithm has been evaluated by the AVIRIS (Airborne Visible/InfraRed Imaging Spectrometer) data provided by NASA's Jet Propulsion Laboratory. A 2.7 times speed up in average can be achieved with four nodes PC cluster, and the detection results is also the same as the serial method running on single CPU.
