【作者】 田潇瑜；

【导师】 徐杨；

【作者基本信息】 中国农业大学 ， 农业工程， 2014， 博士

【摘要】 牛肉作为人类日常消费的一种重要的肉品,因其高蛋白质、低脂肪、维生素及矿物质含量高等特点,受到消费者的青睐。随着人民生活水平的提高,牛肉的品质受到了前所未有的重视。传统检测牛肉品质的方法因其耗时长,效率低,破坏性己不能满足现代化生产的需要,而随着光谱分析技术在农产品检测领域的广泛应用和发展,牛肉品质的快速无损检测技术也成为了研究的热点。本文基于可见近红外光谱和高光谱成像技术,研究了生鲜牛肉嫩度以及相关品质参数的预测模型和评价方法,以实现对牛肉嫩度品质的无损快速检测和评价。具体研究内容及结果如下：(1)对不同光谱系统预测牛肉嫩度进行可行性分析及预测效果评价,得到400-1700nm光谱系统性能更优越,对牛肉嫩度预测准确。研究结果表明光源差异导致不同的信号预处理方式,使用大功率面光源的光谱信号应用变量标准化校正(SNV)后预测模型精度较高。而静态光源系统进行数据平滑(S G)使预测精度改善。两系统都能对牛肉剪切力值进行预测,全波段建模两系统精度相当,筛选特征变量建模时,400-1700nm的光谱系统表现出更优越和稳定的性能。同时考虑系统成本及系统结构以及快速检测需求,最终选择400-1700nm光谱系统为牛肉嫩度最佳检测系统,其预测相关系数和均方根误差为0.9085和7.5212,RPD值为2.16。(2)计算嫩度品质各指标的光谱预处理、最佳预测模型及模型验证,能够实现对各指标较高的预测精度。研究对比了不同预处理方法对各指标的预测效果,并通过建立全波段偏最小二乘回归(PLSR)以及联合区间偏最小二乘回归(si-PLSR)、遗传算法-偏最小二乘回归(GA-PLSR)模型,最后得到剪切力值、a*的最佳预测模型为si-PLSR模型,预测相关系数和标准差分别达到0.9085、0.9027和7.5212、1.4878,模型RPD值为2.16和2.65。L*的最佳预测模型为GA-PLSR,预测相关系数和标准差为0.9457和1.7250,RPD达到了3.24。而对蒸煮损失率的预测效果较差,最佳预测模型为PLSR全波段建模,预测相关系数和标准差为0.8453和2.5054,RPD值为1.84。利用线性判别(LDA)和支持向量机(SVM)对牛肉嫩度等级进行判别,并对牛肉按部位和食用方式确定分级阈值,最后对嫩度等级判别准确率最高达到92.85%。对嫩度品质指标的预测模型和嫩度分级模型分别进行了实验验证,各指标验证相关系数和标准差分别为0.8875、0.9060、0.8972、0.8217和10.16、2.319、1.055、2.493,嫩度判别模型验证预测识别率达到100%。(3)手持式快速检测系统的模型植入和校正,基本实现牛肉品质的在线生产快速检测。根据在线检测的特点和实际检测遇到的问题,对植入的预测模型进行校正,确定了对原始数据进行变量标准化校正(SNV)的预处理最为合理,然后对剪切力值、颜色a*参数使用si-PLSR模型,L*和蒸煮损失率采用PLSR全波建模预测,相关系数和预测均方根误差分别为0.9068、0.9031、0.9049,0.8276和7.1963N、1.8246、1.4931、3.0876,RPD都在2以上,稳定性比较好。用检测系统进行实际生产验证,基本实现了牛肉嫩度品质的在线快速检测,满足了实际生产的需求。并设计开发了软件系统及后台数据库对多模型灵活调用和牛肉嫩度品质多指标实时预测和结果显示,同时可对检测结果进行存储和后期数据查询统计,便于模型的普适性改进。(4)利用高光谱图像分析牛肉特性分布情况,对牛肉样品每一点嫩度分布进行预测。鉴于牛肉样品的结构复杂性和可见近红外光谱单点检测的缺点,本研究利用高光谱成像系统采集牛肉高光谱三维图像,通过逐步回归结合遗传算法(GA)筛选牛肉剪切力值、颜色相关的特征光谱并建立PLSR预测模型,对牛肉高光谱图像进行降维。应用建立的特征波段预测模型,可计算出图像中任一像素点的剪切力值和颜色预测结果,宏观上显示了牛肉样品表面嫩度、颜色特征分布情况；另一方面通过计算特征光谱图像的纹理特征,建立牛肉嫩度等级判别模型并分析牛肉样品嫩度分布情况。更多还原

【Abstract】 Beef, as one of the main meat ware for human daily consumption, is favored by consumers for its high protein, low fat, vitamin and mineral content. With the improvement of people’s living standards, the quality of beef has received the unprecedented attention. The traditional methods for beef detection which are time consuming, low efficiency and destructive already fail to satisfy the need of modern production. With wide use and development of spectral analysis in the field of agricultural products detection, there have been growing interests in the fast nondestructive method for assessing beef quality attributes.Therefore, in this paper, fresh beef tenderness and related quality parameters detecting and evaluating models were researched applying visible near infrared spectral and hyperspectral imaging technology in order to realize rapid nondestructive prediction of beef tenderness. The specific research contents and results are as follows:(1) Research on different spectral system for beef tenderness predicting, analysed the feasibility and evaluate the prediction effect. The difference between two systems, having different light source and resolution, and the results of beef tenderness prediction were discussed. The results indicated that original signal requires different preprocessing methods due to different light structure and characteristic. The spectral using of high-power lights need standard normal variates (SNV) method leaded to high precision for model. For other system with high resolution and low SNR spectrograph, the signal with Savitzky-Golay smoothing (S_G) had improved the accuracy of model. Both the two system can predict beef tenderness well enough using full range wave, the400-1700nm spectral system showed superior and stable performance based wavelength selecting methods, the prediction accuracy of which is higher than dual channel spectral system. Considering the cost, system structure and requirements for fast detection simultaneously, the system in range of400-1700nm were finally chosen for beef tenderness prediction, the correlation coefficient and root mean square prediction error is0.9085and7.5212and RPD value of2.16.(2) The study of pre-treatment methods, prediction models and model validation of fresh beef tenderness, L*, α*and cooking loss. The effects on PLSR prediction models of different preprocessing methods such as savitzky-golay smoothing (SG), multiplication scatter correction (MSC), standard normalized variate (SNV) and first derivative (FD) were studied. The best pretreatment method for tenderness and cooking loss are SNV+SG methods; the best method for L*is MSC+SG; and the original spectral for α*prediction model is best. Then effective variable for each parameters were selected with interval partial least-squares (iPLS), genetic algorithm (GA) to establish synergy interval partial least square regression model (si-PLSR), genetic algorithm-partial least squares regression (GA-PLSR) model. The results showed that the best prediction model for shear-force value, α*were si-PLSR model, with the correlation coefficient and standard deviation0.9085,0.9027and7.5212,1.4878and model RPD value of2.16and2.65, respectively. The best prediction model for L*was GA-PLSR, with correlation coefficient and standard deviation o of prediction set of f0.9457and1.7250, RPD is3.24. The prediction effect for cooking loss is poorer, the optimal prediction model was PLSR full range wavelength, prediction correlation coefficient and standard deviation of0.8453and2.5054, RPD value of1.84. Linear Discriminant Analysis (LDA) and support vector machine (SVM) methods were applied for beef tender level classification. The grading threshold was set as45N and60N according to the beef parts and eating mode, results showed that the LDA is better than SVM model, two types of beef classification accuracy of prediction set were92.85%and91.66%, respectively. The prediction and classification models of all parameters were verified, respectively. The correlation coefficient and standard deviation in validation set were0.8875,0.9060,0.8972,0.8217and10.16,2.319,1.055,2.493, respectively. The beef tenderness classification accuracy of validation set was up to100%.(3) Implanting and correction of models for fast detection system were studied. According to the characteristics of on-line detection and actual problems, prediction model after implanting need correction, the raw data with standardized correction variable (SNV) were most reasonable, and then si-PLSR model for shear force value, color a*, L*and cooking loss using PLSR model full range wavelength, correlation coefficient and the root mean square prediction error were0.9068,0.9031,0.9049,0.8276and7.1963N,1.8246,1.4931,1.8246, RPD all above2. Detection system was verified finally, which basically achieved rapid detection of beef tenderness, have met production demands. Detection software system and background database were designed and developed, Improved the detection software system, and establish the system of background database. Based on which, invocation of the model was flexible and real-time prediction and results display for beef quality parameters, and at the same time prediction results can be carried out for storage and data query statistics, which improved the universality of prediction models.(4) In view of the complexity of beef structure of and the disadvantages of visible near infrared spectral like single point detection, this study used hyperspectral imaging system to get three-dimensional images and more information of samples. The stepwise regression and GA of the resulting PLSR models were used to identify the most important wavelengths and to reduce the high dimensionality of the hyperspectral data. By using these important wavelengths, beef images were developed to predict shear force value and color of every pixel in the images for visualizing in all portions of the sample. On the other hand, texture features were extracted from beef sample images for classification model, and then tender level of beef was calculated in every pixel of the images for visualizing distribution overall the sample.更多还原