CAJ | 학술논문

采用高光谱成像技术（420～1000 nm ）对山楂的缺陷（表面的损伤以及虫害区域）进行识别研究。共采摘了134个样品，包含损伤果46个、虫害果30个、损伤及虫害果10个和完好果48个。考虑到山楂的花萼、果梗与损伤、虫害的RGB图像有相似的外观特征，容易造成误判，利用高光谱成像系统采集了损伤、虫害、完好、花萼和果梗五个区域一共230个山楂样本的高光谱图像，并提取相应的感兴趣区域（region of in‐terest ，ROI），得到了样本的光谱数据。使用标准归一化（standard normalized variate ，SNV ），卷积平滑（savitzky golay ，SG），中值滤波（median filter ，MF），多元散射校正（multiplicative scatter correction ，MSC）方法进行光谱预处理，建立偏最小二乘（partial least squares method ，PLS）判别分析模型，结果表明经过SNV预处理后的预测结果较好。最后选取SNV作为预处理方法。应用回归系数法（regression coefficients ， RCs）从全波段中提取10条特征波段（483，563，645，671，686，722，777，819，837和942 nm ），利用Kennard‐Stone算法将各类样本按照3：1的比例随机分成训练集（173个）和测试集（57个），并对其建立最小二乘支持向量机（least squares‐support vector machine ，LS‐SVM ）判别模型，山楂缺陷的正确识别率为91.23％。然后，运用主成分分析（principal componentanalysis ，PCA）进行10条敏感波段下单波段图像的数据压缩，分别采用“sobel”算子和区域生长算法“Regiongrow”识别出86个缺陷山楂样本的边缘与缺陷特征区域，得出单损伤、单虫害和损伤及虫害样本的识别率分别为95.65％，86.67％和100％。研究结果表明：采用高光谱成像技术可以对山楂的损伤、虫害、花萼和果梗进行定性分析和特征识别，该研究为山楂的缺陷无损检测提供了理论参考。採用高光譜成像技術（420～1000 nm ）對山楂的缺陷（錶麵的損傷以及蟲害區域）進行識彆研究。共採摘瞭134箇樣品，包含損傷果46箇、蟲害果30箇、損傷及蟲害果10箇和完好果48箇。攷慮到山楂的花萼、果梗與損傷、蟲害的RGB圖像有相似的外觀特徵，容易造成誤判，利用高光譜成像繫統採集瞭損傷、蟲害、完好、花萼和果梗五箇區域一共230箇山楂樣本的高光譜圖像，併提取相應的感興趣區域（region of in‐terest ，ROI），得到瞭樣本的光譜數據。使用標準歸一化（standard normalized variate ，SNV ），捲積平滑（savitzky golay ，SG），中值濾波（median filter ，MF），多元散射校正（multiplicative scatter correction ，MSC）方法進行光譜預處理，建立偏最小二乘（partial least squares method ，PLS）判彆分析模型，結果錶明經過SNV預處理後的預測結果較好。最後選取SNV作為預處理方法。應用迴歸繫數法（regression coefficients ， RCs）從全波段中提取10條特徵波段（483，563，645，671，686，722，777，819，837和942 nm ），利用Kennard‐Stone算法將各類樣本按照3：1的比例隨機分成訓練集（173箇）和測試集（57箇），併對其建立最小二乘支持嚮量機（least squares‐support vector machine ，LS‐SVM ）判彆模型，山楂缺陷的正確識彆率為91.23％。然後，運用主成分分析（principal componentanalysis ，PCA）進行10條敏感波段下單波段圖像的數據壓縮，分彆採用“sobel”算子和區域生長算法“Regiongrow”識彆齣86箇缺陷山楂樣本的邊緣與缺陷特徵區域，得齣單損傷、單蟲害和損傷及蟲害樣本的識彆率分彆為95.65％，86.67％和100％。研究結果錶明：採用高光譜成像技術可以對山楂的損傷、蟲害、花萼和果梗進行定性分析和特徵識彆，該研究為山楂的缺陷無損檢測提供瞭理論參攷。
채용고광보성상기술（420～1000 nm ）대산사적결함（표면적손상이급충해구역）진행식별연구。공채적료134개양품，포함손상과46개、충해과30개、손상급충해과10개화완호과48개。고필도산사적화악、과경여손상、충해적RGB도상유상사적외관특정，용역조성오판，이용고광보성상계통채집료손상、충해、완호、화악화과경오개구역일공230개산사양본적고광보도상，병제취상응적감흥취구역（region of in‐terest ，ROI），득도료양본적광보수거。사용표준귀일화（standard normalized variate ，SNV ），권적평활（savitzky golay ，SG），중치려파（median filter ，MF），다원산사교정（multiplicative scatter correction ，MSC）방법진행광보예처리，건립편최소이승（partial least squares method ，PLS）판별분석모형，결과표명경과SNV예처리후적예측결과교호。최후선취SNV작위예처리방법。응용회귀계수법（regression coefficients ， RCs）종전파단중제취10조특정파단（483，563，645，671，686，722，777，819，837화942 nm ），이용Kennard‐Stone산법장각류양본안조3：1적비례수궤분성훈련집（173개）화측시집（57개），병대기건립최소이승지지향량궤（least squares‐support vector machine ，LS‐SVM ）판별모형，산사결함적정학식별솔위91.23％。연후，운용주성분분석（principal componentanalysis ，PCA）진행10조민감파단하단파단도상적수거압축，분별채용“sobel”산자화구역생장산법“Regiongrow”식별출86개결함산사양본적변연여결함특정구역，득출단손상、단충해화손상급충해양본적식별솔분별위95.65％，86.67％화100％。연구결과표명：채용고광보성상기술가이대산사적손상、충해、화악화과경진행정성분석화특정식별，해연구위산사적결함무손검측제공료이론삼고。
Hyperspectral imaging technology covered the range of 380~1 000 nm was employed to detect defects (bruise and in‐sect damage) of hawthorn fruit .A total of 134 samples were collected ,which included damage fruit of 46 ,pest fruit of 30 ,in‐jure and pest fruit of 10 and intact fruit of 48 .Because calyx?s-1 tem‐end and bruise/insect damage regions offered a similar ap‐pearance characteristic in RGB images ,which could produce easily confusion between them .Hence ,five types of defects inclu‐ding bruise ,insect damage ,sound ,calyx ,and stem‐end were collected from 230 hawthorn fruits .After acquiring hyperspectral images of hawthorn fruits ,the spectral data were extracted from region of interest(ROI) .Then ,several pretreatment methods of standard normalized variate (SNV) ,savitzky golay (SG) ,median filter(MF) and multiplicative scatter correction (MSC) were used and partial least squares method(PLS) model was carried out to obtain the better performance .Accordingly to their results ,SNV pretreatment methods assessed by PLS was viewed as best pretreatment method .Lastly ,SNV was chosen as the pretreatment method .Spectral features of five different regions were combined with Regression coefficients(RCs) of partial least squares‐discriminant analysis (PLS‐DA) model was used to identify the important wavelengths and ten wavebands at 483 ,563 , 645 ,671 ,686 ,722 ,777 ,819 ,837 and 942 nm were selected from all of the wavebands .Using Kennard‐Stone algorithm ,all kinds of samples were randomly divided into training set (173) and test set (57) according to the proportion of 3∶1 .And then , least squares‐support vector machine (LS‐SVM ) discriminate model was established by using the selected wavebands .The re‐sults showed that the discriminate accuracy of the method was 91.23% .In the other hand ,images at ten important wavebands were executed to Principal component analysis (PCA) .Using “Sobel” operator and region growing algrorithm “Regiongrow” , the edge and defect feature of 86 Hawthorn could be recognized .Lastly ,the detect precision of bruised ,insect damage and two‐defect samples is 95.65% ,86.67% and 100% ,respectively .This investigation demonstrated that hyperspectral imaging tech‐nology could detect the defects of bruise ,insect damage ,calyx ,and stem‐end in haw thorn fruit in qualitative analysis and feature detection .which provided a theoretical reference for the defects nondestructive detection of hawthorn fruit .