生鲜猪肉中细菌总数(TVC)超标会直接危害大众, 为此研究验证高光谱成像技术结合相应的建模方法预测生鲜猪肉中TVC的可行性。 针对非线性、 小样本问题, 以及光谱维和空间维的大数据量问题, 在综合比较偏最小二乘回归(PLSR)、 人工神经网络(ANNs)和最小二乘支持向量机(LS-SVM)3种建模方法的基础上, 最终选取了LS-SVM方法组建模型。 3种建模方法综合比较的结果表明, LS-SVM同时兼顾了训练精度和泛化能力两方面的性能, 使其都能做到最优, 与标准平板菌落计数法所检测TVC的决定系数分别为0.987 2和0.942 6, 校正均方根误差和预测标准均方根误差分别为0.207 1和0.217 6, 其建模性能优于其他方法。 研究结果表明, 高光谱成像技术结合LS-SVM预测建模方法可作为快速、 非破坏预测生鲜猪肉TVC的有效手段。

Once the total viable count (TVC) of bacteria in fresh pork meat exceeds a certain number, it will become pathogenic bacteria. The present paper is to explore the feasibility of hyperspectral imaging technology combined with relevant modeling method for the prediction of TVC in fresh pork meat. For the certain kind of problem that has remarkable nonlinear characteristic and contains few samples, as well as the problem that has large amount of data used to express the information of spectrum and space dimension, it is crucial to choose a logical modeling method in order to achieve good prediction result. Based on the comparative result of partial least-squares regression (PLSR), artificial neural networks (ANNs) and least square support vector machines (LS-SVM), the authors found that the PLSR method was helpless for nonlinear regression problem, and the ANNs method couldn’t get approving prediction result for few samples problem, however the prediction models based on LS-SVM can give attention to the little training error and the favorable generalization ability as soon as possible, and can make them well synchronously. Therefore LS-SVM was adopted as the modeling method to predict the TVC of pork meat. Then the TVC prediction model was constructed using all the 512 wavelength data acquired by the hyperspectral imaging system. The determination coefficient between the TVC obtained with the standard plate count for bacterial colonies method and the LS-SVM prediction result was 0.987 2 and 0.942 6 for the samples of calibration set and prediction set respectively, also the root mean square error of calibration (RMSEC) and the root mean square error of prediction (RMSEP) was 0.207 1 and 0.217 6 individually, and the result was considerably better than that of MLR, PLSR and ANNs method. This research demonstrates that using the hyperspectral imaging system coupled with the LS-SVM modeling method is a valid means for quick and nondestructive determination of TVC of pork meat.