【作者】 余心杰；

【导师】 刘康生；

【作者基本信息】 浙江大学 ， 运筹学与控制论， 2015， 博士

【摘要】 模式识别是农产品无损检测技术应用及高精度农产品无损检测仪器设备研制的核心环节。本论文针对农产品无损检测中各类原始采集数据的非线性降维和高效建模等模式识别问题,研究建立了基于自编码网络的近红外光谱非线性降维方法、基于稀疏表示的图像和光谱分类识别方法和基于相关向量机的近红外光谱定量建模方法,并将相关方法应用于各类农产品无损检测实践,为农产品无损检测的中模式识别方法研究和应用提供一些新思路和新途径。论文主要研究内容和结果如下：1、基于自编码网络的近红外光谱非线性降维方法研究。针对农产品无损检测中近红外光谱的非线性降维问题,提出了一种基于自编码网络(AN)的近红外光谱非线性降维方法。研究结果表明：采用AN分别对毛竹笋近红外和中红外光谱数据进行流形学习非线性降维提取低维本征信息之后,结合偏最小二乘法(PLS)建立毛竹笋不溶性膳食纤维含量定量回归模型。与其它常用的光谱数据预处理方法(MSC、SNV、Savitzky-Golay)结合PLS回归模型、PCA降维方法结合LS-SVM回归模型,及用单独PLS回归模型预测的结果相比,该方法对毛竹笋不溶性膳食纤维含量预测精度最高,其中NIR光谱的AN-PLS模型预测均方根误差RMSEP为0.0138；MIR光谱的AN-PLS模型预测均方根误差RMSEP为0.0135。可见,AN能有效反映红外光谱中存在的非线性结构,提高了检测精度,为基于近红外光谱技术的农产品检测数据降维研究提供有效的新途径。2、基于稀疏表示的图像和光谱分类识别方法研究。为简化农产品无损检测定性建模方法在参数优化选择、学习训练方面的步骤,提出了一种基于稀疏表示(SR)的图像和近红外光谱分类方法。稀疏表示方法把分类识别问题,转化为一个求解待识别测试样本对于整体训练样本矩阵的稀疏表示问题,只需要在求解L-1最小化范数时简单设定最小误差和迭代次数,就可以完成分类识别任务,有效简化了操作步骤。利用该方法对葡萄干品质机器视觉分类和大西洋鲑鱼肉色近红外光谱分类的结果表明。(1)在对葡萄干品质图像分类识别上,稀疏表示方法比最小二乘支持向量机方法(LS-SVM)取得了接近、甚至更好的分类效果,从而验证了其在农产品无损检测图像分类建模应用中的有效性,为农产品无损检测图像分类建模提供了有效的新途径；(2)在大西洋鲑鱼肉色等级可见／近红外光谱分类建模中,稀疏表示方法对于两种鲜活大西洋鲑肉色等级分类平均准确率为73%,优于线性判别分析(LDA)和最小二乘支持向量机方法(LS-SVM)的分类结果(分别为72%和68%),验证了该方法在农产品无损检测光谱分类建模应用中的有效性,为大西洋鲑肉色活体检测提供了一种新途径,同时也为其它农产品无损检测近红外光谱定性建模提供了一种高效、简便、且具有实用性的建模方法参考。3、基于相关向量机的近红外光谱定量建模方法研究。针对农产品无损检测尤其是在线检测对定量建模方法的预测速度性能要求高的特点,提出了一种基于相关向量机方法的近红外光谱定量建模方法。相关向量机方法无需估计正规化参数,解的稀疏性高,能在保证精度的同时提高预测速度。利用该方法对发酵冬虫夏草菌粉腺苷含量检测,以9000-4000cm-’光谱的主成分分析降维所获得的10个主成分为特征变量输入,建立发酵冬虫夏草菌粉腺苷含量的RVM回归模型,并与PLS、LS-SVM回归模型性能进行比较分析,结果表明：RVM回归模型的预测能力最强。在RVM建模过程中,需要调节的参数只有一个,具有计算性能优势,适用于发酵冬虫夏草菌粉质量近红外光谱在线监测,为农产品无损检测定量建模提供了一种有效的新途径。更多还原

【Abstract】 Pattern recognition is critical for nondestructive detection of agricultural products and development of high precision agricultural nondestructive detection equipment. In this paper, focusing on pattern recognition problems such as nonlinear data dimension reduction and efficient modeling in the nondestructive detection of agricultural products, three pattern recognition methods including a near infrared spectroscopy dimension reduction method based on autoencoder network manifold learning, an image and spectral classification modeling method based on sparse representation, and a near infrared spectroscopy quantitative modeling method based on relevance vector machine, were researched and applied to nondestructive detection of agricultural products. The main research contents and results are as follows:1. Research on near infrared spectroscopy dimension reduction method based on autoencoder network manifold learning. Autoencoder network (AN) is a nonlinear dimension reduction manifold learning algorithm which can find out nonlinear low-dimensional manifold structure from high dimensional spectra data effectively. In this research, a nonlinear infrared (IR) spectra modeling method AN-PLS was proposed by combining AN and partial least squares (PLS) to reflect the nonlinear correlations existing between IR spectra and physicochemical properties of samples. In AN-PLS, AN and PLS were adopted to deduct the dimensions of IR spectra and build regression calibration model, respectively. The AN-PLS was then applied to correlate the near infrared (NIR) spectra and the mid infrared (MIR) spectra with the concentrations of insoluble dietary fiber in bamboo shoots. The results indicate that AN-PLS can predict the concentrations of insoluble dietary fiber in bamboo shoots with a lower cross validation RMS error (RMSECV) and higher determinative coefficient (R2), than other common spectra data preprocessing methods (such as MSC, SNV, Savitzky-Golay and PCA) combined with PLS or sole PLS. It can be concluded that AN-PLS can effectively model the nonlinear correlations between IR spectra and physicochemical properties of the samples. And it is feasible to accurately detect the concentrations of insoluble dietary fiber in the bamboo shoots by coupling NIR and MIR spectra with AN-PLS modeling method.2. Research on image and spectral identification method based on sparse representation. An identification method based on sparse representation was proposed for nondestructive detection of agricultural products using image and spectral technology. In sparse representation, the identification problem was reduced to the problem as how to represent the testing samples from the data dictionary (training samples data). The identification result thus could be achieved by solving the L-1 norm-based optimization problem. We compared the effectiveness of the proposed method with linear discriminant analysis (LDA) and least squares support vector machine (LS-SVM) by appling it to classification of raisin qulity and recognition of Atlantic salmon flesh color. Experimental results demonstrated that (1) the overall identification accuracy of the proposed method for four raisin qulity varieties images was significantly higher than that of LS-SVM, (2) the identification accuracy of the proposed method for two Atlantic salmon flesh color varieties spectrals was 73%, which was higher than that of LDA (72%) or LS-SVM (68%). Therefore, the proposed method can provide a new effective method for identification modeling for nondestructive detection of agricultural products using image or spectral technology.3. Research on near infrared spectra modeling method based on relevance vector machine. A modeling method based on relevance vector machine (RVM) was proposed for nondestructive detection of agricultural products using spectral technology. The relevance vector machine method which with a high sparse solution and without estimating the regularization parameters, can improve the prediction speed. We compared the effectiveness of RVM method with partial least squares (PLS) and least squares support vector machine (LS-SVM) by appling it to analysis adenosine content in fermentation cordyceps powder. Experimental results demonstrated that prediction accuracy of the proposed method for adenosine content in fermentation cordyceps powder was significantly higher than that of PLS and LS-SVM.更多还原