高光谱特征的人造肉中低色度差异物检测

Detection of Low Chromaticity Difference Foreign Matters in Soy Protein Meat Based on Hyperspectral Imaging Technology

人造植物肉在其原料运输、制糜和包装等加工环节时有发生异物污染事件,误食异物会严重损害人的身体健康。常规食品异物检测方法容易检测出如金属、石头等坚硬、深色异物,而软质、浅色、透明异物却是食品异物污染事件中的主要来源且是检测的难点。根据异物和人造肉各自化学组成成分的差异,提出了一种人造肉中低色度差异物的高光谱成像检测方法,根据异物与人造肉光谱信息的差异,建立模式识别模型,来进行人造肉中低色度差异物的判别,最后结合数字图像处理技术对异物进行空间分布可视化。选取了聚碳酸酯(PC)、涤纶树脂(PET)、聚氯乙烯(PVC)、硅胶、玻璃五种食品生产加工过程中常见的低色度差异物为研究对象,模拟人造肉压片的工业制作流程,将异物混入人造肉肉糜中,制备混有异物的人造肉样品,分别采集异物和人造肉感兴趣区域(ROI)的反射高光谱数据,采用SG,SNVT,MSC,VN,1ST及2ND六种不同的光谱预处理方法对原始光谱数据进行预处理,然后采用主成分分析法(PCA)对预处理后的光谱数据降维,采用连续投影算法(SPA)提取人造肉的特征波长。分别以全波段光谱、特征波长和主成分变量作为模式识别模型输入变量,对比LDA,KNN,BP-ANN,LS-SVM四种模式识别模型的准确率,优选出最佳的定性识别模型,设置优选模型异物类别输出变量为1、人造肉类别为0,生成二值图像,再结合数字图像处理技术实现人造肉中异物分布可视化,进而实现人造肉中低色度差异物的识别。结果表明,采用SG预处理后的光谱在降噪方面优于其他预处理方式。SPA法优选了人造肉10个特征波长。全波段主成分变量结合BP-ANN模型的检测效果最佳,准确率达98.33%。验证了高光谱技术应用于人造肉中低色度差异物检测的可行性。

Incidents of foreign matter contamination in the processing links of soy protein meat occur frequently.Consumers’ accidental ingestion of foreign matters will seriously damage human health.Conventional foreign matter detection methods can easily detect hard and dark foreign matters such as metals and stones.Therefore,soft,light-colored,and transparent foreign matters have become the main source of foreign matters in food foreign body contamination incidents and are difficult to detect.Based on the inconsistency of the chemical composition of the foreign matter and the soy protein meat,this study proposes a hyperspectral imaging detection method for the low-contrast foreign matter in the soy protein meat.According to the difference in the spectral information of the foreign matter and the soy protein meat,a pattern recognition model was established to perform soy protein meat and finally combined with digital image processing technology to visualize the spatial distribution of foreign objects.Five kinds of low-contrast foreign matters:polycarbonate(PC),polyester resin(PET),polyvinyl chloride(PVC),silica gel,and glass were selected as the foreign matter in this study.Collecting foreign matter and soy protein meat region of interest(ROI) reflectance hyperspectral data,using SG,SNVT,MSC,VN,1 ST and 2 ND six different spectral preprocessing methods to preprocess the original spectral data,and then use principal component analysis(PCA) to reduce the dimension of the preprocessed spectral data,and use successive projections algorithm(SPA) to extract soy protein meat Characteristic wavelength.Using the raw spectrum,characteristic wavelength and principal component variables as the input variables of the pattern recognition model,try to compare the accuracy of the four pattern recognition models:LDA,KNN,BP-ANN,and LS-SVM,and select the best qualitative recognition model.Set the output variable of the foreign matter category of the optimal model to 1,the category of soy protein meat is 0,generate a binary image,and then combine the digital image processing technology to realize the visualization of the low-contrast foreign matter distribution in the soy protein meat,to realize the recognition of the low-contrast foreign matter in the soy protein meat.The results show that the spectrum after SG pretreatment is better than other pretreatment methods in noise reduction.The SPA method optimized 10 characteristic wavelengths of soy protein meat.The detection effect of the whole band principal component variables combined with the BP-ANN model is the best,with an accuracy rate of 98.33%.