Application of Image Processing and Computer Cluster in the Development of Full-Field Measurement for Electronic Speckle Pattern Interferometry and Digital Image Correlation

本論文應用平行計算技術於非破壞、非接觸檢測之實驗力學，結合數值計算與實驗量測，著重於提升量測訊號與干涉影像之訊噪比，並提升訊號影像後處理之計算速度。於數值計算部份，本文提出一套向量轉化運算流程，將多維度迴圈轉換為等價之二維向量運算，以提升區域性均值濾波、短時傅立葉轉換、區域性最小平方法之運算效能，並與平行運算架構與硬體相互結合，優化全域式量測系統之後處理效能；於實驗量測部份，本文使用高速、高解析之數位攝影機與量測儀器，自行建置全域式量測系統，其中包含電子斑點干涉術、數位影像相關法與光學檢測系統，並應用此架構量測物體之受力微變形與透鏡品質，並進一步應用於自動掃頻系統與自動光學檢測。 電子斑點干涉術部份，本文探討並分析壓電薄板之三維動態特性，並配合阻抗分析儀、雷射都卜勒振動儀、有限元素法，分析其面內外共振頻與共振模態，同時提出一套濾波理論，抑制空氣擾動對干涉條紋之影響，並應用於干涉術之自動掃頻系統。數位影像相關法部份，本文使用高速攝影機、拉伸試驗機、光學顯微鏡，量測物件於控制條件下之動態響應、塑性變形、奈米級變形，同時與光纖位移計相互比較，並獲得一致的量測結果。光學檢測系統部份，本文採用極化光學架構搭配影像處理技術，並提出能量中心檢測邏輯，量測鏡片之非軸對稱傾斜與偏心誤差，能夠快速並廣泛地應用於塑膠射出鏡片與玻璃模造鏡片。本研究成果結合數值計算與實驗量測，並應用影像處理及叢集電腦於電子斑點干涉術及數位影像相關法，同時開發全場分析技術，並與實驗量測、數值計算、與理論解析皆達到相當優異的一致性，該成果可實際應用於學術研究領域或工業界之跨尺度檢測，配合非接觸檢測與光測力學，俾提供即時且直觀之檢測方法。This dissertation applies parallel computation technique into experimental mechanics, and mainly contributes to a) enhancing the signal-to-noise ratio of experimental signal and image, and b) improving computational speed of the denoising algorithms. This study measures an object deformation or profile using three non-contact inspection methods: electronic speckle pattern interferometry (ESPI), digital image correlation (DIC), and lens inspection system. The three methods are full-field measurement and use high-resolution image to record interference fringe, characteristic region, or intensity pattern. An embedded program controls frequency step and loading steps, and a digital camera takes sequential images at each controlled condition, storing the experimental data into a solid disk drive. Analytical programs load thousands of image and billions of pixels, and take several hours to retrieve a denoising result with excellent quality. Computer cluster of central processing unit (CPU) and graphic processing unit (GPU) are used in the computation, and we discuss the performance of denoising algorithm based on CPU and GPU architectures. This article also proposes a vectorizing process which converts for and while loops to equivalent matrix operations, and presents the speedup for different cases. The vectorizing method provides significant gain in analyzing experimental data, in which the displacement filed and strain distribution obtained using DIC method gauge a deformed object on sub-micron scale, while provides high accurate result. The frequency-sweeping curve obtained using ESPI not only indicates the resonate frequency of a vibrating piezoelectric thin plate, but also provides mode shape while approaching and leaving eigenmode, and gives a consistent result compared with those obtained using finite element method, impedance analyzer, and laser Doppler vibrometery