| 摘要: | 由於電腦科技的進步及影像擷取設備的普及，自動光學檢測 (automatic optical inspection, AOI) 在各產業的應用越臻成熟。而且，產業結構的改變，人力資源成本逐年增加，許多關於自動光學檢測的議題越來越受重視，期待用自動化的方式來取代人眼做檢查的工作，並加快檢查速度、增加檢出正確率。相關議題包括生醫影像診斷、去氧核醣核酸電泳分析、蛋白質電泳分析、車輛安全監測、光電元件產品檢測、半導體製程檢測、薄膜液晶體顯示器檢測、紋理分割、人臉偵測、居家監控…等。在本論文中，我們將探討三個與自動化光學檢測相關的議題：去氧核醣核酸電泳分析、薄膜液晶面板瑕疵檢測、及紋理分割。這些議題的前景物件都出現在不均勻的背景中。對於去氧核醣核酸電泳分析，可以使用一維曲線匹配背景；對於薄膜液晶面板瑕疵檢測，可以使用二維平面匹配背景；對於紋理分割，則在超平面 (hyper plane) 匹配背景紋理的賈伯強度 (Gabor magnitude)。這些主題在匹配背景之前，各別需要一些適合各議題影像特性的前處理。
在去氧核醣核酸電泳分析這個主題中，我們提出一個完全自動化的系統，用於檢測脈衝場膠體電泳 (pulsed-field gel electrophoresis, PFGE) 影像內的帶狀條紋 (band)。帶狀條紋針測包含泳道切割與條紋賦予兩部分。泳道切割演算法運用脈衝場膠體電泳影像的特徵並使用最佳化的平行直線匹配方式來分割各泳道。條紋賦予演算法使用多項式匹配來移除不均勻的背景，然後使用條紋的斜率特徵來偵測條紋。
在薄膜液晶面板瑕疵檢測這個主題中，我們針對低對比高雜訊的斑痕瑕疵 (mura) 提出一個線上偵測斑痕瑕疵的方法，包含亮度校正、多重影像累加、多重解析度背景相減、及瑕疵偵測。首先，在移動產線上的薄膜液晶面板從不同角度拍攝許多影像，然後用一張標準影像來校正這些拍攝影像中的不均勻環境亮度。第二，各影像依薄膜液晶面板的位置對齊，並累加相同位置的像素灰階。第三，將累加影像以離散小波轉換得到三階多重解析度影像並估計各解析度的背景，然後從粗糙到細緻逐步修飾所估計的背景。累加影像減去估計的背景後就留下可能是瑕疵的部分。最後，採用一個標準常態分佈灰階的二值化方法來找出薄膜液晶體顯示器的斑痕瑕疵。
在紋理分割這個主題中，我們提出一個非監督式的紋理分割方法，該方法在主動輪廓模式 (active contour model) 裡結合了最佳化非對稱賈伯濾波器 (Gabor filter)。首先，我們提出非對稱高斯函數的模式，然後把該函數乘上一個二維複數三角函數以建構出二維的非對稱賈伯濾波器。然後，計算一張紋理影像的賈伯能量強度並求強度的整體平均與變異數以取得賈伯強度的機率分佈。平均值和變異數用於演化主動輪廓的能量泛函數 (functional)。為了求得輪廓在目前演化狀態下的最佳非對稱賈伯濾波器，我們提出一個擬費雪 (Fisher-like) 函數以決定每次輪廓演化時的最佳非對稱賈伯濾波器。最後，我們陳述完整的主動輪廓演化演算法。而在光電元件瑕疵檢測上，本紋理分割技術將可應用於紋理背景的紋理瑕疵區塊檢測。
在各研究主題的實驗中，我們展示所提自動光學檢測技術的成果，包含：條紋偵測、斑痕瑕疵偵測、及非監督式紋理分割的結果。條紋偵測系統能自動地分割電泳影像的泳道，並於各泳道偵測出條紋，正確率為98.42%。斑痕瑕疵偵測方法能偵測各種方向、形狀、大小的斑痕瑕疵，瑕疵區域能100% 被偵測出來。非監督式紋理分割能區分二個不同的紋理區域，無須預先選擇合適的賈伯濾波器；最後我們也將此技術應用於布匹紋理瑕疵檢測上。
In recent years, object detection has become more popular for industry applications due to the usage of advanced scanning devices and the requirement of visual inspector. Moreover, due to the growth of image and video data, many issues of automatic object detection are expected such as substitution for human inspection, acceleration of inspection speed, and increases on inspection correctness. These issues include biomedical image diagnosis, deoxyribonucleic acid (DNA) electrophoresis analysis, protein electrophoresis analysis, vehicle safety monitoring, solar cell production inspection, semi-conductor wafer inspection, thin-film-transistor liquid-crystal display (TFT-LCD) inspection, texture segmentation, human face detection, house surveillance, etc. In this dissertation, we’ll discuss three of these interesting auto-detection issues: DNA electrophoresis analysis, TFT-LCD inspection, and texture segmentation. Among the three issues, foreground objects appear on non-uniform background. For DNA electrophoresis analysis, backgrounds are fitted by one dimensional curves
for TFT-LCD inspection, backgrounds are fitted by two dimensional planes
for texture segmentation, Gabor magnitudes of texture backgrounds are fitted by hyper planes. Before background fitting, each issue needs some pre-processing which is suitable for the image features of each issue.
In DNA electrophoresis analysis, we proposed a completely automatic band detection system for pulsed-field gel electrophoresis (PFGE) images. Band detection comprises lane segmentation and band assignment. The lane segmentation algorithm characterizes features of the PFGE images and uses optimal line fitting to separate lanes. The band assignment algorithm uses polynomial fitting to remove the uneven background and uses gradient features of bands to detect bands.
In TFT-LCD inspection, we proposed an online TFT-LCD mura defect detection method which consists of illumination calibration, multi-image accumulation, and multi-resolution background subtraction. First, an LCD on a moving product conveyer is contiguously captured by several images with different locations and a synthesized LCD image is used to calibrate the non-uniform illumination of the images. Second, the images are aligned in position to accumulate the gray levels of pixels which all correspond to a point on the LCD. Third, the multi-resolution backgrounds of the accumulated image are progressively estimated based on the discrete wavelet transform (DWT). We take the accumulated image into a multi-resolution and then refine the estimated background from coarse to fine. The accumulated image subtracted from the estimated background leaves the defect candidates. Finally, a standard thresholding method is used to “threshold out” the mura defects.
In texture segmentation, we proposed an unsupervised texture segmentation method using optimal asymmetric Gabor filter (AGF) based on active contour model. First, we create a formula of the asymmetric Gaussian function and multiply a two dimensional (2D) complex sinusoidal function to the function to construct a 2D AGF. Then, compute the average and the variation of the Gabor magnitudes to capture the probability distribution of the Gabor magnitudes. The average and variation are used in the level-set energy functional to evolve the level-set contour. To obtain an AGF which is optimal to the current evolution contour, we propose a Fisher-like function which determines the optimal AGF for the processed image at every iteration determined. Finally, the proposed algorithm of active contour is described.
Experiments demonstrate the proposed automatic object detection techniques: band detection system, mura detection method, and unsupervised texture segmentation. The band detection system can automatically segment the lanes in the gel images and detect the bands in the lanes. The band detection rate is 98.42%. The mura detection method can detect mura defects with arbitrary directions, shapes, and sizes. The detection rate of mura regions is 100%. The proposed unsupervised texture segmentation method can distinguish two different textural regions without pre-selecting a suitable Gabor filter. |
