雙穩態膽固醇液晶顯示器(Bi-stable Cholesteric Liquid Crystal Display, BS-ChLCD)已經被視為具有發展潛力的軟性顯示器(Flexible Display, FD)。然而,正常的雙穩態膽固醇液晶顯示器其面板影像不僅含有週期性的水平紋路,而且具有許多隨機分佈的顆粒狀添加物,使得檢測雙穩態膽固醇液晶顯示器的問題是棘手的。因此,本論文提出了ㄧ個新穎的瑕疵檢測方法,此方法能自動從複雜的背景中檢測出瑕疵。首先,將一張待檢測影像切割成許多小尺寸的子影像。為了得到良好的瑕疵偵測率,本研究採用離散餘弦轉換(Discrete Cosine Transform, DCT)與紋理(Texture)的表示方法從每張子影像中抽取一系列有用的特徵向量。本論文提出了ㄧ種功能強大的機器學習方法當作瑕疵偵測器,稱為支持向量資料描述(Support Vector Data Description, SVDD)。支持向量資料描述能夠判斷輸入向量屬於正常或瑕疵類別。為了使基於支持向量資料描述的瑕疵偵測器具有自我學習的能力,我們事前必須收集一組正常的子影像去訓練此瑕疵偵測器,因此,提出了基於相關性檢驗之訓練樣本自動篩選演算法。藉由結合訓練樣本自動篩選演算法至SVDD中,即發展了一個自組織之SVDD瑕疵偵測器。此策略經過應用在真實的BS-ChLCD面板影像中,其實驗結果顯示本論文所提出的檢測方法其瑕疵偵測率可高達99%。
Bi-stable cholesteric liquid crystal display (BS-ChLCD) has been considered a promising flexible display (FD). However, the defect inspection problem for BS-ChLCD is intractable because the surface of a normal BS-ChLCD contains not only horizontal lines that appear periodically, but also many particle-like additives that randomly distribute over the surface. This paper proposes a novel defect inspection system, which is able to detect defects from complex background automatically. First, an input image is segmented into small-sized sub-images. To obtain good defect detection rate, this paper adopts discrete cosine transform (DCT) and textural representation methods to extract a set of useful feature vector from each sub-image. In this paper, a powerful machine learning method called support vector data description (SVDD) is adopted as the defect detector. SVDD is able to judge whether an input vector belongs to normal class or defect class. However, we have to collect a set of normal sub-images to train the SVDD-based defect detector beforehand. In order to let the detector have the self-learning ability, a correlation checking-based automatic training sample selection (CC-ATSE) algorithm is also proposed. By introducing the CC-ATSE algorithm into the SVDD, a self-organizing SVDD (SO-SVDD) defect detector is developed. Experimental results, carried out on real BS-ChLCD images, show that the proposed SO-SVDD-based defect detection scheme can achieve a high defect detection rate over 99%.