鏡面玻璃製品已成為生活必需品或建築及電子產品重要材料之ㄧ,而表面瑕疵的存在會直接影響鏡面玻璃表面反射出的影像品質,因此表面瑕疵的偵測對鏡面玻璃之品質檢驗相當重要。鏡面玻璃在生產過程中,因在透明玻璃背面進行金屬濺鍍後產生鏡化之特性提高鏡面玻璃之表面反射能力,所以在人工檢驗時常容易受到表面反射外界物體影像之干擾造成瑕疵誤判,加上鏡面玻璃表面容易附著灰塵、水漬等,使用傳統影像處理亦容易將此異物誤判為瑕疵,不易辨別瑕疵存在之位置。因此本研究針對鏡面玻璃製品,發展一套適用於拍攝車用鏡面玻璃表面影像之機台及瑕疵檢測系統,進行玻璃表面瑕疵之自動化檢測。 本研究提出以各不重疊區塊為單位進行區塊離散餘弦轉換(Block Discrete Cosine Transform, BDCT),並擷取轉換後之頻率係數絕對值和位置相關之能量集中係數計算出頻率係數之衰減率,利用灰預測(Grey prediction)中非線性柏努力灰預測模型(Nonlinear Grey Bernoulli Model, NGBM)針對衰減率進行建模,並重新修正衰減率,以有效衰減區塊中頻率係數值,達到去除背景中的干擾且能突顯瑕疵之目的。實驗結果顯示本研究所提方法之瑕疵檢測率達96.45%,瑕疵誤判率為0.0674%,優於傳統頻率域之瑕疵檢測方法。
Mirror-glass products have become necessities in our daily life and major materials for construction and electronic industries. Since the surface defects directly affect the quality of the mirror-glass products, the detection of surface defect is very important for manufacturers. In the production process of mirror-glass products, a metal is platted on the back of transparent glass to improve the property of reflection. Human inspection is easy to be interfered by the external object images reflected on the surface of mirror glass and results in making erroneous judgments of defect detections. Moreover, the surface of mirror-glass product is easily attached to dust, dirt, water, and so on and make the defect inspection tasks more difficult. Therefore, this research aims at exploring the automated inspection of surface defects of car mirror glasses. This study proposes using independent blocks as units to do Block Discrete Cosine Transform (BDCT), and calculating decrement rates from absolute values of BDCT frequency coefficients and energy concentration factors. The prediction method of Nonlinear Gray Bernoulli Model (NGBM) in gray theory is used to modify the decrement rates. Then, the BDCT frequency coefficients are declined based on the modified decrement rates and transformed back to spatial domain. Finally, a threshold value can be easily found to segment the surface defects from the background. Experimental results show that the defect detection rates achieve up to 96.45% and the false alarm rates lower to 0.0674% by the proposed method and outperform the traditional defect detection methods.