台灣目前為了確保國人食用雞隻的安全,防檢局在全台各合格屠宰場均設有多名獸醫師負責檢查雞隻,故本研究利用本實驗室所開發之超光譜影像系統分析雞隻屠體光譜以減少獸醫師之負擔以及方便業者針對各種不合格雞隻病灶追溯上游養雞場的疾病,達到協助雞場管理。本研究利用線掃瞄方式拍攝合格與不合格雞隻屠體胸部背部位置,擷取雞胸的9個位置的線光譜來進行光譜處理與分類,合格雞隻由主軸分析(PCA)後的不同Score值配合LDA進行分類,分類結果以胸部訓練組40個Score的誤判率9.0%及測試組的9.3%為最佳,且Score值越多誤判率會降低。不合格雞隻病灶則是利用PCA第一至第三Loading的比值配合 LS-SVM進行分類,特定特徵分類結果以胸部的訓練組誤判率0.75%及測試組的2.76%為最佳;在不同位置分布的分類的結果,以胸部的取樣點2、5、8(中橫)為最佳,訓練組的誤判率為0.75%,測試組的誤判率為3.01%,其重要的波段集中在450nm與800nm左右,綜合特徵以近紅外光的二次微分10、20點距10、20分數時分類結果較佳。其中胸腹水、炎症滲出物與其他病灶間的測試組分類結果誤判率為5.13%、3.85%、9.07%、7.58%,這些結果未來可應用於不合格雞隻的病灶分類。
To ensure the safety of chicken meal, the BAPHIQ (Bureau of Animal and Plant Health Inspection and Quarantine) deploys more than 400 veterinarians in slaughter houses to inspect each chicken carcass in Taiwan. The inspection job is a tedious and time consuming which needs to be improved with automation technique. This study applied a home-make spectral imaging system to detect the carcasses and to identify their lesions along with to classify their wholesomeness. Line spectra were obtained from breast portion and back portion. Nine ROIs of each portion were analyzed by PCA (principle component analysis) scores and loadings. Results show classification error is 9.0% and 9.3% for training and testing data set, respectively, when 40 scores were used in model. More score used in model deceased the classification error. Nine ROIs suggested that the middle portion of ROI region (label 2, 5, and 8) have the lowest classification error. Test results also indicates that when the ratio of the first to third loading of PCA were used in model along with differential treatments on spectrum, LS-SVM approach could have error on lesion classification of 5.13, 3.85, 9.07 and 7.58% on four major lesions: abnormal, ascites, inflammatory extrudes, and the others. Loading analysis also suggested important bands for identification locating around 450 nm and 800 nm. These findings might assist on further development for automatic machine.