應用影像處理於皮革不可用面積之自動測量

Abstract

皮革的廣泛應用與皮革依可用面積作為交易而不斷產生紛爭，使得皮革製造業者對於皮面瑕疵的檢驗日趨急迫，傳統的人工檢測非但缺乏標準性、無法自動化，且其因瑕疵所產生的不可用面積計算，也常因檢測人員的個人因素而有所誤差，因此本研究的主要目的，便在於利用影像處理自動依標準作業偵測出皮面瑕疵，並將近距離瑕疵做群組，計算出真正不可用的皮面面積。 研究中第一部份是收集皮面瑕疵，並將其依形狀分類與偵測，總計可分為圓點、細點、線狀、條狀、不規則、圖案、破洞等七類。考量色度與亮度為檢測之重要因素，故實驗上對每一種瑕疵均以黑白和彩色分別檢驗之。又考量彩色影像由RGB三原色組成，皮面瑕疵顏色甚多，對RGB三原色的敏銳不同，故在彩色影像上分別對R、G、B做檢測後，再將三原色檢測結果予以綜整，得到最終之檢測結果。本研究在二值化閥值求取方式上均採用OTSU【1979】所發展方式與Matrox公司在Inspector上所建立的閥值計算法兩種，以比較其異同，結果顯示，七種形狀中，只有破洞有較佳的檢測結果，其他形狀瑕疵因與背景的色差較小，若以本研究方法作檢測則效果較不理想，必須再使用其他影像處理技巧逐一修整後方可進行近距離瑕疵群組作業。 第二部分是近距離瑕疵群組，本研究發展出一套水平、垂直、對角的群組演算法，將影像分別自0度~44度做旋轉及群組，再將群組後的影像回轉及整合，即可完成群組作業，結果顯示，對各形狀之群組結果與實務上依人工所計算的專家面積相較，誤差均在+-10%以內，成效良好。

In the first part of the study, we categorize the defects according to their shapes. There are seven categories：spots、thin spots、strips、irregulars、patterns、and holes. Because color and brightness are two important factors that will alter inspection results, so we take monochromatic and color images of each leather for experiments. For color images which are composed of RGB colors, every image may differ their reaction to color image processing. So three processing were used to process R、G、B colors separately and these three processings were integrated into one final result. The binarization method used in this study follows that of OTSU【1979】and Matrox Inspector. The comparison about these two binarization methods is also implemented. Only holes hold better inspection results in this study. Other inspection results of different shapes are not satisfactory by using the method mentioned in this study and need other post processing to obtain better inspection results. The second part of the study is to develop a defect-grouping algorithm which processes the leather image in vertical、horizontal、and diagonal directions. The procedure must have the image to be rotated from 0 to 44 degrees and have the defects to grouped, and have the image to be rotated back to original position. Finally the unrotated and integrated image is returned to complete the grouping task. The errors of final results are within ±10% and are similar to those made by human experts suggests that the proposed method is feasible for calculating the unusable leather surface.