車用後視鏡是交通工具中不可或缺的必備物件,透過反射車旁與後方路況,提供駕駛相關資訊以減少不必要之交通意外產生,常扮演行車安全中一關鍵角色。車用後視鏡生產過程中,若烘曲、電鍍、凹折及磨邊等製程出現異常,易造成車用後視鏡刮傷、粒點、針孔及缺角之4類瑕疵的產生,將嚴重影響鏡面玻璃表面反射出的影像品質,進而提高行車危險性。目前生產線檢驗大多以人工目視居多,長時間檢測下易造成漏檢及誤判,並針對尺寸檢驗大小上採取錯誤矯正措施導致輪廓尺寸分類錯誤率上升,所以本研究發展一套適用於拍攝車用後視鏡玻璃瑕疵之檢測系統,自動檢測可疑瑕疵,以降低不良率。 本研究主要針對瑕疵檢測與尺寸大小檢驗兩部份進行探討,表面瑕疵方面,提出利用二維離散傅立葉轉換搭配高通濾波處理後配合凸殼演算法檢測出瑕疵。輪廓瑕疵本研究則提出空間域的影像處理方法,使瑕疵位置得以突顯,並同樣運用凸殼演算法處理影像形狀,將鏡面中的雜訊去除分離出瑕疵。而尺寸檢驗使用了指數加權移動平均管制圖執行微量偏移偵測,確認樣本點超出上下管制界限判定異常後,進行EWMA數據與蕭華特管制圖的數據平均值比對以分類出車用後視鏡輪廓尺寸大小。本研究現階段轉換公制之計算可檢測出最小瑕疵大小為0.44mm,提出針對兩種類型瑕疵使用凸殼演算法可發現能有效降低表面瑕疵之正常區域誤判率及提高輪廓瑕疵檢出率。實驗顯示瑕疵檢出率可達90.47%,正常區域誤判率為4.42%,正確判斷率為95.66%,而尺寸大小分類總正確率目前可達85.45%,於檢測與尺寸量測檢驗兩端皆具不錯之效益。
Car mirrors are indispensable essential in object reflection and play a key role in driving safety. In the production process of the car mirrors, some operations such as baking, electroplating, recesses, edging, etc. could be controlled abnormally. This could easily produce scratches, bubbles, pin holes, chips, the common surface and profile defects on car mirrors. These defects will seriously affect the surface quality of the mirror reflection and increase the driving risk. At present inspection of car mirror in production lines, most tasks are conducted by human inspectors. Human inspection is easy to be interfered by the external objects reflected on the surfaces of mirrors and results in making erroneous judgments of defect inspection. Therefore, this research aims at exploring the automated surface defect inspection and dimensional measurement of car mirrors. In defect inspection of car mirrors, we propose using discrete Fourier transform with high-pass filter and convex hull algorithm to detect surface defects. Meanwhile, this study also proposes using the morphology methods for enhancing object contours then applying convex hull algorithm to detect profile defects on mirror images. In dimensional measurement of car mirrors, we propose using the exponentially weighted moving average control method for detecting small shift variation of mirror edge points, and checking sample points beyond or lower the upper and lower control limits. Experimental results show that the defect detection rate achieves up to 90.47%, the false alarm rate is lower 4.42%, and the dimension classification rate is up to 85.45%. These indicate the proposed system is effective in both of the defect detection and dimensional measurement.