本研究是以機器視覺技術來偵測三類不同底色的PVC卡片在品管作業階段時所發生的表面瑕疵。由於在PVC卡片的製造過程中,常因印刷機器之溢墨而產生汙點、印刷面仍殘留異物便進行壓合以及產品運輸或搬運過程而產生的瑕疵,所以本研究主要偵測的瑕疵種類有尺寸不良、流痕、刮痕、汙點、氣泡、異物等。 本研究對於PVC卡片的表面瑕疵偵測主要分成硬體設備的架構與影像處理分析兩大部份。在硬體設備架構部份,因PVC卡片表面屬於高反射體,所以本研究採用暗場打光原理來解決表面反射問題。在影像處理分析部份,分成樣品比對與表面瑕疵偵測,在文字偵測是以影像連結體為主來萃取文字的幾何特徵,且透過歐式距離來做特徵比對;至於圖案則採用兩階段式搜尋方法與三步搜尋法來找出最相似區塊,最後並記錄樣品比對後之位置。在表面瑕疵偵測則是使用影像增強中的直方圖等化來增強PVC卡片上對比性低的瑕疵特徵,並以移動可調式濾波器及個別統計管制界限法,來突顯對比性低的異常點,以達到瑕疵偵測目的。經實驗結果顯示本方法可偵測出低對比的瑕疵,其平均瑕疵檢出率為92.7%,而總平均偵測時間約為0.25~0.31秒。因此,可實際應用於反光面上低對比性的表面瑕疵偵測,以改善人眼檢測所造成的缺失。
In order to solve the problem that defects arise in the quality control step. This study develops a system to inspect PVC card’s surface defects of three kinds of colors by using machine vision. During the manufacturing process of PVC cards, the defects are generated because of the painting machine and artificial neglect. In this study, the types of defect are inaccurate size, flashes, scratches, spots, bubbles and pollution. This study divided into two steps, including hardware establishment and image processing. Because of the high reflecting surface of the PVC cards, we can not acquire image totally and show the defects, and it’s meaningless to other process. Therefore, applies dark-field imaging technique to capture images to solve the problem. It is divided into pattern matching and surface defect detection. In the pattern matching, we applied the Blob method to extract the characteristic of the words, and we used Euclidean distance to match the characteristics, recorded the position of the word. As to the pattern, we take two phase search method and three step search algorithm to find the similar block, record the position after pattern matching. This study take histogram equalization to enhance contrast of surface defect, shift the block adaptive filter and use statistical control limits to intensify the low-contrast abnormal point. Experimental results have shown that the proposed method can effectively detect small defects on low-contrast surface. The average accuracy of inspection is 92.7%, and the average detection time is 0.25~0.31 seconds. In this study, we can improve the drawbacks of human vision inspection on detecting the low-contrast surface defects.