基于高光谱技术的酸奶中常见致病菌的快速鉴别及计数

酸奶是一种发酵型乳制品饮料，因其特殊的功能性和良好的口感而广受欢迎。但由于商业链的不正当运行，如奶源非法获取、灭菌不充分等原因，导致酸奶中致病菌大量滋生，酸奶中毒事件频繁发生。酸奶中常见的致病菌主要有大肠杆菌、金黄色葡萄球菌和沙门氏菌，这三种致病菌由人体摄入并达到一定的数量时会产生腹痛、腹泻等严重的消化道疾病，并且会破坏人体肠道内的正常菌群平衡，因此国标对奶制品中这三种致病菌的数量已有明确的限量规定。由于酸奶的主要消费对象为老人和小孩，故其潜在危害不容小觑。传统菌落检测方法虽具有简单，灵敏、可操作性强等优点，但当不同菌落混杂在一起时无法同时进行定性定量的检测，且具有试剂成本高，检测周期长，人为因素影响较大等缺点。因此开发一种快速、简单、准确的混合鉴定计数方法为避免致病菌对酸奶的潜在危害提供了有效的途径。高光谱技术同时包含样本的光谱信息与图像信息，既能够根据化学组分的微小变化进行精确识别（光谱信息），又能够反映出菌株在外部多层次的变化（图像信息）。因此该研究尝试对比高光谱图像技术和光谱技术，采用模式识别的方法，对比不同的模型识别结果，优选出最佳识别率的识别模型作为计数模型，最后通过最佳鉴别计数模型的识别分类结果来达到对酸奶中常见致病菌鉴定计数的目的。首先，购买酸奶中常见的乳酸菌种（保加利亚乳杆菌、嗜热链球菌、嗜酸乳杆菌、干酪乳杆菌、植物乳杆菌）和潜在污染的致病菌种（金黄色葡萄球菌、大肠杆菌、沙门氏菌）等标准菌株进行培养，提取经过48 h培养后的菌落图像信息和光谱信息。采用几种不同的预处理方式（SNV，MC，MSC，1stDER，2ndDER）对所提取的光谱数据进行预处理，并应用遗传算法筛除光谱数据中冗余的波段，保留有效波段。利用图像处理技术对图像信息中的菌株与培养基背景进行去除，然后采用主成分分析法从每幅图中优选出3个特征波长，并运用图像处理技术从特征波长所对应图像中提取菌株的18个基于GLCM的纹理特征信息。挑选合适的主成分分别建立不同的鉴别模型（LDA，KNN，BP-ANN，LS-SVM），通过其最终的鉴别模型的识别率来确定最佳鉴别计数模型。最后从标准菌株中分别挑选出30株进行计数测试，通过比较模式识别的分类数量结果与菌株的实际数量来验证模式识别效果的准确率。研究表明，运用SNV预处理后光谱数据在提高信噪比效果上明显优于其他几种预处理方式。745.790 8，773.098 4和779.207 0 nm为图像信息中方差贡献率最大的三个波长，运用从特征波长所对应的图像中所提取的纹理特征信息建立图像识别模型。通过对比图像信息和光谱信息的模式识别结果发现，光谱特征鉴别模型普遍优于图像纹理特征鉴别模型，且当主成分数为9时，运用光谱特征所建立的LS-SVM模型的校正集识别率为96.25%，预测集的识别率为91.88%，为最优模型。采用优选的最优模型对菌株进行识别计数，大肠杆菌计数的相对误差为3.33%，金黄色葡萄球菌和沙门氏菌计数的相对误差均为0，验证了高光谱技术应用于酸奶中常见致病菌的鉴别计数的可行性。

Rapid Identification and Enumeration of Common Pathogens in Yogurt Using Hyperspectral Imaging

Yogurt is a kind of fermented dairy beverage, and it is celebrated for its special functionality and good taste. However, due to the improper operation of the commercial chain, such as the illegal acquisition of milk sources and so on, the pathogenic bacteria in yogurt are widespread, resulting in frequent occurrence of yogurt poisoning. The main pathogenic in yogurt are Escherichia coil, Staphylococcus aureus and Salmonella. Human consumption of these three kinds of bacteria will cause severe digestive tract diseases and destroy the balance of normal flora in the human intestine after reaching a certain number. Therefore, the Chinese National Standard has a clear limit on the number of the three pathogens in dairy products. Because the main object of yogurt consumption is the old and the children, the potential harm of yogurt should not be underestimated. The traditional colony detection method is sample, sensitive and operable, but when different colonies are mixed together, it can not be qualitatively detected at the same time, and there are shortcomings such as high cost, long detection cycle and human factors. Therefore, it is of great practical significance to develop a fast, simple and accurate mixed identification count method to avoid the potential hazards of pathogenic bacteria in yogurt. Hyperspectral technology integrates the spectral information and spatial location information of the sample. It can not only accurately identify according to the tiny change of chemical components (spectral information), but also reflect the multi-level changes of the strain (image information). Therefore, this study adopts pattern recognition method to compare different models established by the image information and spectral information, and selects the best counting model based on the recognition rate of the model. Finally, the identification and counting of common pathogenic bacteria in yogurt were realized by the classification results of the best model. Firstly, the standard strains of lactic bacteria and potentially contaminated pathogenic bacteria in yogurt were cultured, and the colony image information and spectral information after 48 h, culture were extracted. Then, different pre-processing methods (SNV, MC, MSC, 1stDER, 2ndDER) were used to reduce the spectral data, and the genetic algorithm was used to reduce the excess spectral bands. The image of agar background used image processing technology to mask removal, then 3 characteristic wavelengths were selected from each map by principal component analysis, and 18 texture feature based on gray-level co-occurrence matrix texture information were extracted from the strain of characteristic wavelengths. Different discriminant models (LDA, KNN, BP-ANN, LS-SVM) were established by selecting the appropriate principal component, and the best discriminant model was determined by the recognition rate of the final discriminant model. Finally, 30 strains from each standard strain were selected for counting test, and the accuracy of pattern recognition was verified by comparing the results of classification and quantity of pattern recognition with the actual number of strains. The results showed that the spectral data pretreated by SNV were superior to other pre-treating methods. The 745.790 8, 773.098 4 and 779.207 0 nm were the characteristic wavelengths. Through the contrast of image pattern recognition and spectral information rate results, it was found that the spectral characteristics of the differential model were better than those of the image texture feature identification model, and when the number of principal component was 9, the LS-SVM spectral model was the optimal model, and the recognition rate of the correction set is 96.25%, and the recognition rate of the prediction set is 91.88%. The optimal model was applied to recognize and count the strains. The relative error of Escherichia coil count was 3.33%, and the relative error of count of Staphylococcus aureus and Salmonella was 0, which verified the feasibility of applying hyperspectral technology to identify and count common pathogenic bacteria in yogurt.