Lin Hao, Zhao Jie-Wen, Chen Quan-Sheng, Cai Jian-Rong, Zhou Ping
School of Food and Biological Engineering, Jiangsu University, Zhenjiang 212013, China.
Guang Pu Xue Yu Guang Pu Fen Xi. 2010 Apr;30(4):929-32.
Near infrared (NIR) spectroscopy combined with pattern recognition was attempted to discriminate the freshness of eggs. The algorithm of one-class support vector machine (OC-SVM) was employed to solve the classification problem due to imbalanced number of training samples. In this work, 86 samples of eggs (71 samples of fresh eggs and 15 samples of unfresh eggs) were surveyed by Fourier transform NIR spectroscopy. Firstly, original spectra of eggs in the wave-number range of 10 000-4 000 cm(-1) were acquired. And then, principal component analysis (PCA) was employed to extract useful information from original spectral data, and the number of PCs was optimized. Finally, OC-SVM was performed to calibrate discrimination model, and the optimal PCs were used as the input eigenvectors of model. In order to obtain a good performance, the regularization parameter v and parameter sigma of the kernel function in OC-SVM model were optimized in building model. The optimal OC-SVM model was obtained with nu = 0.5 and sigma2 = 20.3. Experimental result shows that OC-SVM got better performance than conventional two-class SVM model under the same condition. The OC-SVM model was achieved with identification rates of 80 for both fresh eggs and unfresh eggs in the independent prediction set. The identification rates of fresh eggs were 100% in two-class SVM model. However, when the two-class SVM model was used to discriminate the unfresh eggs of, the identification rates were 0% in the independent prediction set. Compared with conventional two-class SVM model, the OC-SVM model showed its superior performance in discrimination of minority unfresh eggs samples. This work shows that it is feasible to identify egg freshness using NIR spectroscopy, and OC-SVM is an excellent choice in solving the problem of imbalanced number of samples in training set.
尝试运用近红外（NIR）光谱结合模式识别技术来鉴别鸡蛋的新鲜度。由于训练样本数量不均衡，采用一类支持向量机（OC-SVM）算法来解决分类问题。在本研究中，利用傅里叶变换近红外光谱对86个鸡蛋样本（71个新鲜鸡蛋样本和15个不新鲜鸡蛋样本）进行了检测。首先，采集了鸡蛋在波数范围为10000 - 4000 cm⁻¹ 的原始光谱。然后，采用主成分分析（PCA）从原始光谱数据中提取有用信息，并对主成分数量进行了优化。最后，运用OC-SVM进行判别模型的校准，并将优化后的主成分作为模型的输入特征向量。为了获得良好的性能，在构建模型时对OC-SVM模型中的正则化参数v和核函数参数σ进行了优化。得到的最优OC-SVM模型为ν = 0.5且σ² = 20.3。实验结果表明，在相同条件下，OC-SVM比传统的二类支持向量机模型性能更优。在独立预测集中，OC-SVM模型对新鲜鸡蛋和不新鲜鸡蛋的识别率均为80%。在二类支持向量机模型中，新鲜鸡蛋的识别率为100%。然而，当使用二类支持向量机模型鉴别不新鲜鸡蛋时，在独立预测集中识别率为0%。与传统的二类支持向量机模型相比，OC-SVM模型在鉴别少数不新鲜鸡蛋样本方面表现出卓越的性能。本研究表明，利用近红外光谱鉴别鸡蛋新鲜度是可行的，且OC-SVM是解决训练集中样本数量不均衡问题的一个优秀选择。
