在現階段工業檢測趨勢,已漸漸的由機器視覺於自動化檢測取代傳統人工檢測。在過去,若以人工檢測會因訓練不足、素質不一及疲勞等因素而造成人工檢測的缺失,進而導致無法判斷出瑕疵的問題,甚至還會影響整個生產品質。因此應用機器視覺在工業上不但能提高生產效率、使產品品質達到標準,更能顯現出其價值與重要性。傳統樣板比對方法使用於自動瑕疵檢測時,必須建立參考影像並儲存於資料庫中,再以建立的參考影像與待測影像做比對,以達到瑕疵檢測的目的。因此,傳統樣板比對法的缺點除了耗時之外,且需要較大的儲存空間來存放參考影像。除此之外,使用傳統樣板比對法進行影像比對時,也容易受到環境之影響,如光源、位移及旋轉等,而造成檢測的誤判。 本研究所探討BGA基板之邊緣輪廓的基本組成圖形主要為直線、圓弧與特殊凸形輪廓之規律組合,因此本研究基於去除基本組成圖形即可凸顯異常瑕疵的概念。承述前說明,並基於BGA板上基本組成圖形的設計條件,如直線邊緣之斜率與圓弧邊緣之半徑等為已知的條件下,利用霍氏轉換(Hough transform)及相關度指標來偵測這些基本組成圖形,並配合為去除規律性圖形所設計的一些幾何限制規則(Constraint rules),來消除待測影像中具「規則性」的基本組成圖形,以凸顯不規律的異常區塊。本研究對於106張BGA基板影像檢測結果顯示,BGA基板影像中共約有317個瑕疵(平均每張約有3個瑕疵)的瑕疵檢測效果約有93.4%的瑕疵檢出率,而漏檢率為6.6%(型II誤差)及平均每張影像的誤檢瑕疵數為0.23個(型I誤差)。
In this research, a non-referential machine vision method is proposed to detect surface defects on ball grid array (BGA) substrates. Traditional automatic visual inspection systems have used template matching for PCB inspection. They require a large amount of memory storage, and suffer from environmental changes such as alignment, process variations, lighting, etc. Since the layout of BGA substrates is generally composed of basic primitives such as line segments, circular arcs, and other elements of regular shapes, the proposed method uses the Hough transform, correlation measures and geometric constraints to eliminate all regular primitives on BGA substrates. The remaining elements on the resulting image are those that have irregular shapes, and can be easily identified as defects. With the proposed approach, no reference template is required. It provides an efficient and flexible approach for automatic BGA defect inspection. Experimental results on more than 100 BGA substrate samples have shown the efficacy of the proposed method.