【作者】 杨波；

【导师】 吕小毅；

【作者基本信息】 新疆大学 ， 信息与通信工程， 2022， 硕士

【摘要】 近红外光谱可以反映出有机物中含氢官能团的含量,已经在食品掺假检测和地缘鉴别中得到了广泛的使用。但光谱技术在食品分类鉴别中会受到光谱收集成本的限制,导致训练数据少,而且得到的一维光谱数据的维度也很高。此外,现有研究中在光谱收集、特征提取和使用数据的维度上也比较单一。因此,本研究以新疆特色食品孜然、小茴香和葡萄干作为数据基础,针对以上这些不足,分别进行了以下三部分实验:1.根据光谱鉴别方面遇到的训练数据少,数据维度过高的问题,构建了生成对抗网络结合近红光谱分别对新疆、甘肃和山东三地孜然和小茴香的地缘检测模型。实验中分别对比了传统机器学习算法中基于主成分分析的二次判别模型、基于主成分分析的多层感知机模型和深度学习中改进的AlexNet模型和生成对抗网络模型的预测结果。此外,该实验不仅分别完成了对不同地域的孜然和小茴香的地缘检测,还在保持模型结构和超参数不变的情况下,对同一地域的孜然和小茴香进行了辨识。实验结果显示,生成对抗模型以相互竞争、共同学习的方式不仅取得最好的分类性能,还可以克服训练数据少对模型的限制,减少数据降维的处理流程。2.在数据预处理方式单一的问题上,论文中提出了多形态特征融合策略实现对有限光谱数据进一步的特征挖掘。实验中,以新疆吐鲁番的红香妃、马奶提和木纳格三种葡萄干表皮和果肉的红外光谱作为数据基础,对比了仅使用果皮光谱数据、果肉光谱数据、二者数据直接求平均和对二者光谱进行特征融合这四种方式下的实验结果。构建的线性分类模型和非线性分类模型的结果均表明采用果皮和果肉特征融合的实验方式会比其他数据处理方式更好。这表明多形态特征融合可以为光谱数据采集和特征融合这两方面提供新的策略。3.在解决光谱特征提取和处理数据维度单一的问题上,以上述的孜然、小茴香和葡萄干数据为基础,讨论了多维数据分别在提高数据维度PCANet和降低数据维度堆叠式稀疏自动编码器这两种特征提取方式下对模型性能的影响。实验中分别比较了传统的机器学习算法中的K最邻近算法和支持向量机与深度学习中的全卷机神经网络和多尺度卷积网络四个模型在多种评价指标下的结果。各种指标显示,两种特征提取方式在二维光谱数据的形式下更有潜力大幅度提升模型性能,这也为以后食品光谱在实验数据形式处理和特征提取方法的选择上提供了参考。更多还原

【Abstract】 Near-infrared(NIR)spectroscopy can reflect the content of hydrogen-containing functional groups in organic compounds,and has been widely used in food adulteration detection and geographical identification.But spectral techniques in food discrimination strategies will be limited by the cost of spectral collection,resulting in little training data,and the spectral data dimension is also high.In addition,the pretreatment method,feature extraction method and use data of the acquisition spectra in the experiment are also relatively single.Therefore,this study took Cumin,Fennel and Raisin as the data basis of Xinjiang specialty foods,and carried out the following three parts of the experiment for these deficiencies:1.According to the phenomenon of less training data and high dimensionality of spectral data encountered in spectral identification,a generative adversarial network(GAN)combined with NIR spectroscopy was built to a geo-detection model for cumin and fennel in Xinjiang,Gansu,and Shandong province.In the experiment,the results of quadratic discriminant analysis based on principal component analysis(PCA-QDA),multilayer perceptron based on principal component analysis(PCA-MLP)in traditional machine learning algorithms,and the improved AlexNet and generative adversarial network(GAN)in deep learning algorithms were compared respectively.In addition,the experiment not only completes the traceable detection of cumin and fennel in different regions but also discriminates cumin and fennel in the same region with the models structure and parameters remained unchanged.The results of multiple evaluation indexes show that GAN not only achieves the best classification performance by competing with each other and learning together but also can effectively overcome the limitation of little training data and reduce the processing of spectral dimension reduction.2.On the problem of single pretreatment method of collected spectral data,this paper proposes a morphological feature fusion strategy to realize further feature mining of limited spectral data.In this experiment,the NIR spectra of three kinds of raisin skin and flesh,Hongxiangfei,Manaiti and Munage of Turpan in Xinjiang,were used as the data set to compare the experimental results of these four ways using only skin spectral data,flesh spectral data,both data were directly averaged and the two spectra features were fused.The results of the constructed linear classification models and nonlinear classification models show better experimental results with skin and flesh feature fusion than with other data processing ways.It also shows that multi-morphological feature fusion can provide a new strategy for spectral data acquisition and feature fusion.3.On the basis of cumin,fennel and raisin’s spectral data,this paper discusses the impact of two feature extraction methods that improving the data dimension PCANet and reducing the data dimension Stacked Sparse Automatic Encoder(SSAE)with multidimensional data on the performance of the models.In the experiment,the results of K-Nearest Neighbor(KNN)and Support Vector Machine(SVM)in traditional machine learning algorithms and Fully Convolutional Networks(FCN)and multi-scale convolution network in deep learning algorithms are compared under various model evaluation indexes.Various indicators show that the two feature extraction methods have more potential to greatly improve the performance of the models in the form of two-dimensional spectral data,which also provides a reference for the experimental data processing and feature extraction methods of recording food spectra in the future.更多还原