查詢結果分析
來源資料
相關文獻
- 應用高光譜影像於星蘋果表面缺陷偵測與分析
- Investigating the Use of Spectral Angle Mapper and Support Vector Machine for Surface Defect Detection of Jujubes
- 可攜式高光譜影像儀應用於遙測空氣品質指標
- 鋁片冷軋用油之特性需求及對鋁片表面品質之影響
- 高光譜資料之光譜差異分析及量化指標
- SPOT綠度與森林生育地因子於臺灣杉材積推估之探討
- Estimation of Rice Growth from Reflectance Spectra of Vegetative Cover
- 植生指數在集水區內的地文分佈研究
- 偵測鋼胚表面溫度變化的表面缺陷檢驗
- 高光譜影像特徵萃取方法之探討
第1筆 /總和 1 筆
/ 1 筆
頁籤選單縮合
題名 | 應用高光譜影像於星蘋果表面缺陷偵測與分析=Using Hyperspectral Images for Skin Defect Detection of Star Apple |
---|---|
作者 | 阮吳明智; 劉乃上; 范國天; 周浩平; | 書刊名 | 農業工程學報 |
卷期 | 69:4 2023.12[民112.12] |
頁次 | 頁62-73 |
分類號 | 434.251 |
關鍵詞 | 高光譜; 星蘋果; 表面缺陷; 波段比; 植生指數; 光譜角映射法; 支撐向量機; Hyperspectral image; Star apple; Surface defects; Band ratio; Vegetation index; Spectral angle mapper; Support vector machine; |
語文 | 中文(Chinese) |
DOI | 10.29974/JTAE.202312_69(4).0005 |
中文摘要 | 本研究應用波長範圍為450~990 nm的高光譜數據於星蘋果常見表面缺陷(銹斑、疤痕、蟲咬與斑點)的辨識與分類。除了使用波段比與植生指數進行二元分類之外,也使用光譜角映射法與支撐向量機進行多元分類。結果顯示使用波段比進行二元分類時波段配對R730/R618、R730/R685、R723/R640與R755/R570;使用植生指標進行二元分類時,波段配對R685/R570、R730/R685、R685/R640與R685/R570分別對銹斑、疤痕、蟲咬與斑點有最佳的辨識效果。使用450~990nm波長範圍全波段(450波段)為輸入數據建立的光譜角映射法與支撐向量機多元分類模型精度分別為68.9%與96.4%。使用由主成分分析所提取的特徵波長(14個波段)為輸入數據所建立的降維光譜角映射法與支撐向量機多元分類模型精度分別為67.4%與94.7%。與使用全波段的支撐向量機多元分類模型比較,使用特徵波長為輸入的支撐向量機模型表面缺陷分類精度雖然降低了1.7%,但是能大大的降低所使用的波段數目,有潛力被應用於星蘋果採收後處理的自動化。此外本研究發現眩光對於應用高光譜影像資訊於深紫色星蘋果表面缺陷辨識有高度的負面影響,因此建議在對深色且表面具有蠟質的水果建立實用的光學表面缺陷檢測系統時,宜採用間接或漫反射光源以降低眩光的影響。 |
英文摘要 | This study investigated the feasibility of using hyperspectral image data within the wavelength range of 450 ~ 990 nm for surface defect detection of deep-color star apples with some typical surface defects (i.e., Rusty, Scarring, Insect bite, and Spots). The Band Ratio (BR) and Vegetation Index (VI) were used for binary classification; furthermore, Spectral Angle Mapper (SAM) and Support Vector Machine (SVM) were used for multiclass classification of surface defects. Principal Component Analysis (PCA) was performed on hyperspectral image data of star apple samples with typical surface defects to obtain the candidates of optimized wavelength pairs of BR and VI methods and the characteristic wavelengths for data reduction of SAM and SVM models. The results showed that the use of band pairs R730/R618, R730/R685, R723/R640, and R755/R570 for BR; and band pairs R685/R570, R730/R685, R685/R640, and R685/R570 for VI could obtain the optimized binary classification results for Rusty, Scarring, Insect bite, and Spots respectively. Furthermore, for multiclass classification, the accuracies of SAM and SVM models using all hyperspectral image data in the 450 ~ 990 nm range (450 bands) were 68.9 % and 96.4 %; and the accuracies of SAM and SVM models using the characteristic wavelengths (14 bands) obtained by PCA are 67.4 % and 94.7 %, respectively. Compared to the SVM model using complete hyperspectral image data, the overall accuracy of the SVM model using only 14 characteristic wavelengths decreased slightly (1.7 %); however, the gain of wavelength band reduction is vast. Thus, the SVM model using characteristic wavelengths has the potential to be used for sorting deep-color star apples when performing the postharvest process. Furthermore, we found that the glare on the smooth and waxy deep-color star apples severely affects all surface defect detection methods used in this study. Thus, indirect or diffusive light sources are recommended to minimize glare's effect when using RGB or spectral images for surface defect classification of deep-color fruits with waxy surfaces. |
本系統之摘要資訊系依該期刊論文摘要之資訊為主。