基于高光谱成像技术的羊肉新鲜度预测

为实现羊肉新鲜度的快速、无损检测,应用高光谱成像技术对不同存储天数的羊肉建立挥发性盐基氮TVB-N(Total volatile basic nitrogen,TVB-N)含量预测模型。通过高光谱成像系统获取羊肉样本935～2 539nm的高光谱图像,选取样本左上、左下、右上、右下和中间5个位置20×20像素的方形作为感兴趣区域(Region of interesting,ROI),提取ROI并计算区域内样本平均光谱。利用二进制粒子群优化算法(Binary particle swarm optimization,BPSO)和相关系数分析法(Correlation coefficient,CC)提取羊肉TVB-N高光谱特征变量,结合偏最小二乘回归(Partial least squares regression,PLSR)和随机森林回归(Random forest regression,RFR)建模算法,分别建立表征羊肉TVB-N含量的BPSO-PLSR、BPSO-RFR、CC-PLSR、CC-RFR预测模型。依据袋外均方根误差RMSEOOB最小原则,对最佳回归子树和分裂特征2个主要参数进行寻优以提高RFR建模算法的预测精度。比较4个模型的预测效果发现,BPSO-RFR模型的预测精度最高,其校正集决定系数R_c~2和均方根误差RMSEC分别为0.87和2.99,预测集决定系数R_p~2和均方根误差RMSEP分别为0.86和3.36。综上,高光谱成像技术和机器学习算法的有机结合为快速有效检测肉品新鲜度提供了理论依据。

Nondestructive detection of lamb freshness based on hyperspectral imaging technology

In order to achieve rapid and nondestructive detection of lamb freshness, the total volatile basic nitrogen(TVB-N)prediction model was established for lamb with different storage days by hyperspectral imaging technology. Hyperspectral images of lamb samples in the near infrared(935-2 539 nm)regions were acquired by the hyperspectral imaging system. Taking the square of 20×20 pixels in the top left, bottom left, top right, bottom right and middle of the sample as the region of interest, the average spectrum of the sample was extracted and calculated. The hyperspectral characteristic variables of lamb TVB-N content were extracted using the binary particle swarm optimization(BPSO)and correlation coefficient(CC)method. Combining partial least squares regression(PLSR)with random forest regression(RFR)modeling algorithm, BPSO-PLSR, BPSO-RFR, CC-PLSR and CC-RFR prediction models for TVB-N content of lamb were established, respectively. In order to obtain the optimal numbers of regression subtrees and splitting features, the main parameters of the model were traversed and optimized according to the convergence for root mean square error of out-of-bag estimation. Comparing the prediction effects of the 4 models, BPSO-RFR prediction model achieved the highest accuracy. The R2_c and RMSEC were 0. 87 and 2. 99 for calibration set, respectively, and the R2_p and the RMSEP were 0. 86 and 3. 36 for prediction set, respectively. In short, the study demonstrated that hyperspectral imaging technology in combination with machine learning algorithm could effectively detect lamb freshness rapidly and non-destructively, and provide theoretical basis for developing rapid and effective detection method.