數位影像相關法之自動特徵擷取及計算效能提升

Abstract

數位影像相關法(Digital image correlation, DIC)是一種非接觸式全場量測的影像光學量測技術，常應用在跨尺度量測的工程領域及實驗力學的學術研究上。數位影像相關法可以透過影像追蹤待測物上的特徵，將作為特徵的樣板子集合與拍攝的影像序列進行相關係數(Normalized cross correlation, NCC)的計算，再搭配次像素(Sub-pixel)演算法，獲得高精度的位移、速度及應變等多種物理量。 實驗上有時需要量測試片上確切位置的位移或位移，會以圓形斑點作為特徵，而在定義欲追蹤的樣板影像時，經常是以人工方式選取特徵點再獲取樣板影像，但當需要同時追蹤多點特徵時，便需要花費大量的時間選取特徵。為了在各種影像中自動擷取特徵，本文使用影像處理及邊緣偵測的方法加強並擷取所有特徵，再利用大津演算法(Otsu’s method)來過濾並保留位於前景的特徵，最後以特徵周圍的亮度值來篩選特徵，獲取最終的樣板影像。 本文也提出在數位影像相關法中，以計算出的相關係數值為基準來進行樣板影像的更新，以減少反向合成高斯牛頓法(Inverse compositional Gauss-Newton method, ICGN method)迭代運算時所需的迭代次數，同時加入像素選擇策略來減少計算所需的像素數量，先以電腦生成斑點影像測試不同參數設定的精度及計算效率，也藉由拍攝懸臂薄板振動、黏彈性材料的拉伸及鋼珠落擊薄膜等實驗，分別探討穩態振動的條件、特徵具有大變化的準靜態大變形、以及高速衝擊下的暫態變化，驗證處理實際拍攝影像時，此方法對於提升運算速度以及量測準確度的影響，除了能夠量測較大應變的情形，對於DIC系統解析連續影像變化的強韌度亦於本研究有所貢獻。

Digital Image Correlation (DIC) is a non-contact, optical measurement technique for full-field measurement. It is commonly used in engineering for measuring across different scales and in academic research for experimental mechanics. DIC can track features on the test object through image tracking and calculate the correlation coefficient by normalized cross correlation (NCC) between the template image and the captured image sequence of the feature. Combined with sub-pixel algorithms, DIC can obtain high-precision physical quantities such as displacement, velocity, and strain. When experiments require measuring the exact displacement of a specimen, it is often necessary to manually select feature points to define the template image. As tracking multiple feature points, selecting features might be time-consuming. To automatically capture features from various images, image processing and edge detection methods are used to enhance and capture all features. The Otsu method is then used to filter and retain foreground features, and the brightness values around the features are used to select the final template image. This thesis proposes using correlation coefficients to determine whether to update the template subset when calculating the inverse-compositional Gauss-Newton method (ICGN). Additionally, a pixel selection strategy is introduced to decrease the number of pixels involved in the calculations. The improvement in accuracy and speed was tested using computer-generated speckle images with different parameter settings. This method is tested in the experiment by the images from vibration, quasi-static tensile test, and high-speed impact. A cantilevered plate tests the steady-state dynamics. The viscoelastic material is tested by tensile test at various speeds. A high-speed steel ball impacts the large deformation of a polymer membrane. The results show to improve computation speed and measures more enormous strains effectively.