| 摘要: | 隨著技術與科學的快速發展，科學家已開發出許多高解析度的光學顯微鏡，研究者能夠透過高解析度的光學顯微鏡來觀察尺度越來越小的樣品。然而，光學顯微鏡景深的限制，導致拍攝的二維影像只有部分區域的解析度清晰可見，造成影像資訊和觀測品質受到影響，二維影像不佳會帶來三維影像建構不精確與大面積影像縫合困難等問題；另外，傳統光學顯微鏡僅能觀測樣品的二維影像，與三維影像相比它無法讓研究員瞭解和分析樣品之真實高度、立體形狀、表面輪廓等資訊。因此，開發一個新的影像演算法來改善影像品質不佳問題，同時利用影像拼接與重建技術來產生高精確與大範圍的三維影像是當前重要研究的議題之一。
若要求光學顯微鏡拍攝出一張高解析度的二維影像，不可避免需要犧牲可觀測範圍，結果通常無法將整個量測的樣品包含在內。因此，透過控制壓電平台來移動樣品，藉此取得樣品不同位置的二維影像有助於擴展觀測範圍，同時也可以兼顧解析度之需求。不幸的是，壓電平台具有許多非線性特性影響控制追蹤表現。因此，設計一個先進的控制器來大幅減緩壓電平台的非線性特性是另一個重要的議題。簡言之，若要建立出一個高精確與大範圍的三維光學顯微影像系統，一個新的影像演算法和一個先進的控制器的設計不可或缺。
在本論文中，所提出的三維光學顯微影像系統的機構設計可以分為量測部分和掃描部分，在韌體設計包含了三維影像演算法和粒子群最佳化-倒傳遞類神經網路-比例積分微分控制器。論文所提出的三維影像演算法是由影像金字塔轉換融合方法、影像縫合技術、二維到三維轉換方法三個部分所組成。具體而言，新的影像融合規則包含最大局部區域能量法和脈衝耦合神經網路模型，來應用於影像金字塔轉換融合方法。在控制器的設計，我們結合粒子群最佳化、倒傳遞類神經網路、比例積分微分演算法三個部分。透過提出的三維光學顯微影像系統來達成高精確與大範圍的樣品三維影像之目標。所提出的控制方法與現今多種常用的控制方法比較，藉由模擬和實驗結果來證實，所提出的控制器具有更好的追蹤有效性和優越性，因此非常適合用於本論文所建構的三維光學顯微影像系統。在實驗部份，我們利用已知的標準校準樣品來驗證與量化所提出的三維光學顯微影像系統的量測能力。我們利用主觀和客觀兩種評價方式證實利用新的影像融合規則得到的融合影像明顯優於其它典型的影像融合方法。最後，我們將所提出的系統應用於量測未知的微機電系統結構樣品，對於未知結構的樣品量測表現同樣具有優異的三維影像建構能力。;With the rapid evolution of technology and science, many high-resolution optical microscopes (OMs) have been developed by scientists to let researchers be capable of observing smaller and smaller samples. However, the OM’s depth of field (DOF) is limited leading to only partial areas of the captured two-dimensional (2D) images being clear, resulting in information and quality from 2D images being compromised, creating imprecise three-dimensional (3D) images, and difficulties in performing image stitching. Moreover, a high-resolution sample 2D image captured by the OM usually cannot cover the entire measured sample. Also, compared to 3D image microscopic systems, 2D OM image systems cannot allow researchers to understand and analyze the height, shape, and surface profile of the samples. Therefore, developing a novel 3D image algorithm to address the OM’s limited DOF problem, execute image stitching, and produce 3D images is urgently required.
When a high-resolution 2D image is taken by an OM, the observable area inevitably needs to be sacrificed; hence, the entire measured sample is usually not included. Thus, the piezoelectric stage is designed to move the measured sample, which helps increase OM measurement performance. Unfortunately, the piezoelectric stage has unavoidable nonlinear characteristics that will impact the controller’s tracking performance. Consequently, designing an advanced controller to address the piezoelectric stage’s nonlinear characteristics is essential. In a nutshell, it is imperative to establish a 3D optical microscopic imaging (3DOMI) system with a novel 3D image algorithm and an advanced controller.
In this thesis, the proposed 3DOMI system is classified into measurement and scanning parts, including the 3D image algorithm and PSO-BPNN-PID controller. The 3D image algorithm is composed of an image pyramid transform (IPT)-based fusion method, image stitching, and a 2D to 3D conversion approach. Specifically, the novel image fusion rule contains the maximum local area energy method (MLAEM) and pulse-coupled neural network (PCNN) model applied to the IPT-based fusion method. The PSO-BPNN-PID controller combines particle swarm optimization (PSO), back propagation neural network (BPNN), and proportional-integral-derivative (PID) algorithms. The target is to generate extensive and highly precise sample 3D images through the proposed 3DOMI system. Comparing the proposed control method with various traditional control schemes, we used simulations and experimental results to demonstrate that the proposed controller is effective, superior, and more suitable for the proposed 3DOMI system. Furthermore, a known standard sample will be used to calibrate and quantify the measurement capability of the proposed 3DOMI system. In particular, subjective and objective evaluation methods confirm that the fused images obtained by the novel image fusion rules outperform other classical image fusion methods. Ultimately, the proposed 3DOMI system is applied to measure an unknown microelectromechanical systems (MEMS) structure sample, which also performs excellently. |
