- 學位論文
混合模型平行計算運用於遙測影像特徵抽取與分類法上之分析
A Novel Hybrid Model of High Performance Parallel Computing Applied to the Analysis of Feature Extraction and Classification for Remote Sensing Images
摘要
近年來在圖形識別的領域裡，如何從原始的高維資料中，抽取高品質與穩定度的特徵，再根據所選取的特徵透過一個有效的分類器進行分類，進而獲取較好的辨識率，已漸成為研究的主流。然而在遙測領域中在面對大量資料處理時，本論文提出了以特徵為基礎之平行概念的特徵選取方法及以測試樣本為基礎之平行概念的分類器，來解決近年來，已成熟的高維資料量感測技術所帶來的龐大計算問題。 本研究分為兩個部份，第一部分是建立平行近似架構來降低資料維度，再使用平行k-way SM (Semi-Matroid)做資料分類；第二部份是提出混合模型平行架構，利用叢集電腦中每部電腦所設置中央處理器(Central Processing Unit, CPU)和圖形處理器(Graphic Processing Unit, GPU)來完成混合平行系統。 本研究使用的硬體架構是利用MPI(Message Passing Interface)來完成每部電腦之間的資料傳送與接收，並在每部電腦中分別使用OpenMP和CUDA(Compute Unified Device Architecture)的函式庫來處理CPU和GPU的任務分配，建立出一套高計算能力的混合模型平行架構系統。
並列摘要
In the research of Graphic knowledge in recent years has become a main flow as how to extract the characteristics of stabilized and high-quality from the original high-dimensional data, and then classifies it from an effective classification with the characteristics of the selected sub-types to obtain a better identification rate. However, in the field of facing large volume data processing, in this theory will propose based in the characteristics of implementation of feature selection methods as the samples of the parallel classifier in order to solve the problem of high-dimensional data flow sensor technologies arising from the massive calculation. This research will look into two areas, first is to establish the best approximation of the parallel structure to reduce the data dimension, and then use of parallel k-way SM (Semi-Matroid) for the data classification; second is proposed the framework of the parallel hybrid model which use of cluster computer settings for each computer Central Processing Unit (CPU) and Graphic Processing Unit (GPU) to achieve the parallel hybrid system. In this research, the hardware architecture is use Message Passing Interface (MPI) to complete transferring and receiving date between each computer, and each computer has using OpenMP and Compute Unified Device Architecture (CUDA) to library CPU and GPU in order to deal with the distribution of tasks, so it can creates a hybrid model of parallel structure.