| 摘要: | TFT-LCD 液晶顯示器製造廠商為了獲得更好的利潤，除了製作高附加價值的產品之外，製程上也要求成本的降低及報廢的減少，本研究以玻璃切割製程為背景，去探討利用機器視覺在高滲透壓刀輪切割後的玻璃斷面檢測之方法，用以取代現有的人工檢查方式並建立預警機制減少報廢。
檢測方法中的前處理運用邊緣檢測、二值化、型態濾波、細線化，將上述處理完的圖像使用水平投影得到玻璃斷面像素的統計圖，再利用截尾平均值的極端值判定方法去判斷 Rib Mark Line 的存在與否，接著以霍夫轉換及最小平方法兩種不同的直線檢測方式量測 Rib Mark Line 的距離。檢測方法的優劣以影像判斷正確率、平均執行速度及量測準確度為評定標準，量測準確度的評估使用平均絕對值誤差(MAPE)，實驗程式及介面使用Microsoft Visual C++ 2010 和 OpenCV 設計，實驗樣本以厚度 0.5mm 的玻璃取像 95 張無瑕疵影像以及 130 張瑕疵影像，共 225 張做為測試樣本，經由測試及檢測方法的改良，瑕疵和無瑕疵影像判斷正確率 100%，平均執行速度可以達到 0.094 秒，MAPE 值 5.93%，最後再投入 0.7mm 厚度的玻璃無瑕疵影像 30 張及瑕疵影像 30 張，測試結果影像判斷正確率 100%，平均執行速度和 MAPE 值與 0.5mm 厚度玻璃實驗相近，此方法可成功應用在不同玻璃厚度的斷面檢測。
;In order to improve profits for TFT-LCD liquid crystal display manufacturers, other than producing high value-added products, firms can also reduce waste in the manufacturing process itself. This study focuses on the background of the glass cutting process, in order to explore the use of a machine vision in a cutter wheel after high osmotic pressure glass cutting that can replace the existing manual inspection methods. A reduction in waste can also be achieved by the implementation of an early warning system.
The pre-processing of the detection method includes edge detection, binarization, patterns filtering, and thinning, then the image is processed using horizontal projection to achieve a graph of the pixel cross-section of the glass. Afterwards, the truncated average end value determines whether the rib mark line is present. Two different detection methods, the Hough transformation and the least squares method, then measure the amount of straight-line distance of the rib mark line. There are advantages and disadvantages of different image detection methods when determining the correct rate, average execution speed, and measurement accuracy. The accuracy assessment experimental design was performed using the MAPE value, with Microsoft Visual C ++ 2010 running OpenCV. There were 95 flawless images and 130 flawed images taken of the experimental samples which had a glass thickness of 0.5mm. In total, 225 test sample images were taken, and through improved testing and detection methods, the flawed and flawless image correct image determination rate was 100%. The average execution speed was 0.094 seconds, with MAPE values of 5.93%. Finally, there were 30 flawless images and 30 flawed images taken of experimental samples which had a glass thickness of 0.7mm, and the flawed and flawless correct image determination was again 100% correct. The average execution speed the MAPE values of 0.7mm glass thickness samples were similar to the 0.5mm glass thickness experimental results. This method can be successfully applied to determine of different glass cross-sectional thicknesses. |
