摘要
本研究比較與探討應用光譜角映射法與支撐向量機於蜜棗表面缺陷檢測之精確度。為了獲得蜜棗表面的高光譜影像，使用了波長範圍為468-950 nm的推掃式高光譜影像系統對包含銹斑、裂果、發霉、過熟腐爛等特徵的高雄11號(珍蜜)蜜棗進行掃描。應用相關高光譜影像資料建立了光譜角映射法與支撐向量機蜜棗表面缺陷分類模型對蜜棗表面缺陷進行辨識。由於蜜棗有弧度的表面光滑並具有蠟質，進行光譜影像掃描時表面的眩光無法避免，因此在本研究中將表面的眩光視為蜜棗表面的特徵，在建立分類模型時眩光被視為分類的類別之一。由相關模型的辨識結果建立混淆矩陣以評估模型分類有效性。由光譜角映射法與支撐向量機模型對蜜棗表面缺陷分類的混淆矩陣顯示光譜角映射法模型在分類時對於裂果與眩光的混淆度最高。支撐向量機模型對銹斑的辨識度較低，對白色發霉的部分辨識度最高。使用相關分類模型對驗證數據的分類結果顯示光譜角映射法與支撐向量機模型的分類準確性分別為71.5%和97.2%。由此研究可知由波長範圍為468-950 nm的高光譜數據所建立的支撐向量機模型能對蜜棗表面缺陷進行良好的辨識。
關鍵詞：高光譜影像，蜜棗表面缺陷檢測，支撐向量機，光譜角映射法Abstract
This study investigated the classification accuracy of spectral angle mapper (SAM) and support vector machine (SVM) models for surface defect detection of jujubes with the use of hyperspectral image data. Hyperspectral imaging data of "Kaohsiung 11" jujubes with surface defects including rusty spots, decay, white fungus, black fungus, and crack were obtained by using a custom made hyperspectral imaging system. Since the surface of jujubes is smooth and waxy, the glare on the surface cannot be avoided during spectral image scanning. Therefore, in this study, the glare on the surface is regarded as the characteristics of jujubes. Confusion matrices were used to evaluate the effectiveness of models. The classification accuracy of SAM model was 71.5%. For SAM classifier, the confusion between crack and glare is the highest. The classification accuracy of SVM model was 97.2%. For SVM model, the prediction accuracy of rusty spots is the lowest and the accuracy of white fungus is the highest. The high classification accuracy of SVM model indicated that the SVM classifier, with the use of hyperspectral imaging data, is effective for classifying surface defect of jujubes.
Keywords: Hyperspectral Imaging, Jujube Surface Defect Detection, Support Vector Machine, Spectral Angle Mapper