應用螢光高光譜影像檢測混合動物油與植物油之性質

Abstract

近年來發生多起食品安全事件，在經由文獻以及實際訪問後得知油品的混摻是為常見且現今仍然存在的問題且牛油是為較常見的混摻對象之一，傳統的食品檢測流程需要大量人力及時間成本，在此考量之下本研究致力於應用相對快速且非破壞性的高光譜螢光影像於食品性質之檢測方法。本研究以檢測牛油為主要對象，分別混合了米糠油、芥花油以及橄欖油於五種不同濃度 (100%, 75%, 50%, 25%, 0%) ，透過建立各個油品的激發散射矩陣找出油品的激發光波段。並以選定的激發光搭配高光譜影像系統收集光譜數據，分別使用了人工神經網路 (Artificial Neural Network, ANN) 建立混合油種的分類判別模型。預測模型效果上純牛油、米糠油、芥花油以及橄欖油的準確率可達7.33%、99.50%、100.00%、68.25%。另亦使用輔助向量迴歸 (Support Vector Regression, SVR) 方法預測牛油與米糠油、芥花油以及橄欖油的混合油中的牛油濃度平均誤差為14.58%、3.81%、10.64%。以及考量到不同廠牌造成之影響後，使用ANN分類得純牛油、米糠油、芥花油以及橄欖油預測準確率為88.53%、97.4%、82.13%、59.5%。使用SVR方法預測為15.00%、7.61%以及11.31%。使用ANN模型在預測桌上型檢測裝置所量測之數據之準確率，在分類純牛油、摻雜米糠油、摻雜芥花油、摻雜橄欖油的準確率為97.00%、100.00%、100.00%、100.00%。使用SVR模型在預測桌上型檢測裝置所量測之數據之準確率，在預測牛油分別與米糠油、芥花油以及橄欖油的混合中牛油的濃度平均誤差分別為7.18%、10.16%以及11.92%。

In recent years there’s lots of food safety issue around the world, according to the paper review and market survey we conclude that oil adulteration is one of the most commonly event in actual problems, whereas traditional inspection process takes time and laborious so as a respond this research aims to discover the potential of using Hyperspectral fluorescence imaging to assess the oil concentration. In this research, Ghee is the adulteration target, which will be adulterated with Rice bran, Canola, Olive oil in five concentration level (100%, 75%, 50%, 25%, 0%) respectively, first the use of Excitation and emission matrix gives the information about what exact excitation light should be used to induce fluorescence, use this excitation as a light source in Hyperspectral imaging system to get the fluorescence spectrum for each oil mixture, and build the ANN model to classify the adulterated type, for pure Ghee, adulterated Rice bran, Canola, Olive oil the accuracy approaching 7.33%, 99.50%, 100.00%, 68.25% respectively, SVR model for predicting the Ghee concentration in oil mixture reaching averaging error 88.53%, 97.4%, 82.13%, 59.5% respectively, furthermore, if also takes the different brand factor for each plant oil into account, the accuracy for ANN still retain 100%, 96%, 93.5% respectively, for SVR model we have averaging error 15.00%, 7.61%, 11.31% respectively. For ANN classification accuracy of portable detection device also reaching 97.00%, 100.00%, 100.00%, 100.00%, for SVR model also reaching average error 7.18%, 10.16%, 11.92%, respectively.