【作者】 石礼娟；

【导师】 文友先；

【作者基本信息】 华中农业大学 ， 农业机械化工程， 2011， 博士

【摘要】 随着社会经济的发展和人民生活水平的提高,粮食供求的主要矛盾已经从数量的不充足转变为质量的不理想。因此,建立稻米质量快速无损检测方法,无论是在育种,还是在食品加工和农产品贸易中都有重要意义。本文以不确定性人工智能理论和化学计量学理论为基础,利用数字图像处理技术、光谱分析技术、小波分析和模式识别技术研究影响稻米食用品质的三种重要品质垩白、直链淀粉含量和陈化的快速无损检测方法。为了提高机器视觉对垩白的识别精度与适应能力,在稻米垩白品质快速无损检测方面,主要研究了以下内容：(1)构建了用于图像采集的计算机视觉系统。分别在透射光与反射光环境下研究了光源、电压、背景等因素对稻米图像质量的影响,从而确定了适合稻米图像采集的最佳环境条件,即透射光条件下的最佳拍摄条件为：电压水平为6.4V,光源为LED灯组,背景为浅蓝色；反射光条件下的最佳拍摄条件为：电压水平为6.0V,光源为LED灯组,背景为深蓝色。(2)分析了透射光与反射光条件下采集的稻米图像的直方图,研究了合适的图像去噪方法与图像分割算法。通过不同色彩空间下彩色直方图与灰度直方图分析发现反射光下采集的图像的灰度分布适合稻米垩白区域的识别。设计了加权均值滤波模板,既消除了噪声,又保护了胚乳区与垩白区的边界。根据图像分割算法-最大类间方差法的适用范围确定了稻米图像在进行垩白识别时应截取的矩形区域的大小。(3)研究了具有自适应能力的垩白识别算法。该方法以不确定性人工智能理论与云模型为基础,把垩白与非垩白定义为两个定性概念,以一个不对称云和一个对称云来分别表达垩白与非垩白,以两组数字特征分别描述垩白云与非垩白云。通过对比同一电压下人工目测法、固定阈值法,云分类法的垩白大小检测结果检验云分类法的精确度,试验结果表明,云分类法比人工目测法的精确度高,云分类法与固定阈值法(即准确值)的偏差的均值为0.97,人工目测法与固定阈值法(即准确值)的偏差的均值为1.93；通过对比不同电压下人工目测法、云分类法的垩白大小检测结果检验云分类法的适应性,试验结果表明,云分类法比人工目测法的适应性好,用云分类法计算不同电压下同一粒米的垩白大小的标准差均值为0.57,人工目测法为2.29。为了建立稳定性好、预测精度高的稻米直链淀粉含量近红外光谱定量分析模型,在稻米直链淀粉含量快速检测方面,主要研究了以下内容：(1)研究了光谱采集参数对稻米近红外光谱响应特性的影响。通过对同一直链淀粉含量的稻米在不同参数下采集的光谱的统计分析,确定最佳采集参数为：扫描次数为64,分辨率为8cm-1,室内温度为15℃。(2)研究了稻米异常光谱剔除方法和光谱预处理方法。为了优化校正集,提高模型的预测精度,用基于马氏距离准则和基于预测浓度残差准则相结合的方法剔除了18条因各种主客观因素产生的异常光谱。为了消除由于基线的漂移与偏移、仪器的随机噪声、杂散光等对光谱产生干扰,提高光谱的信噪比,采用多种方法对稻米光谱进行预处理,并比较了这些方法对建模结果的影响。通过模型评价指标的比较,确定用原光谱经一阶导数与SG卷积滤波相结合的方法进行光谱预处理。(3)研究了定量分析方法对建模效果的影响。分别用逐步多元线性回归、主成分回归和偏最小二乘回归三种定量分析方法对经导数与SG卷积滤波处理后光谱建立校正模型。比较各项模型评价指标发现,偏最小二乘回归(PLS)方法建立的校正模型稳定性最好,预测值与标准值的相关系数最高,预测均方差最小。预测值与标准值相关系数为98.96%,校正均方差为0.62,预测均方差为1.19,交叉检验均方差为1.58。为了建立陈化稻米近红外光谱定性识别模型,本文主要研究了以下内容：(1)研究了陈化米和非陈化米的近红外光谱响应特性,并利用主成分分析法结合马氏距离研究了不同光谱预处理方法对聚类效果的影响,一方面确定了用稻米的近红外光谱进行定性识别的可行性；另一方面,综合考虑类内距离小,类间距离大的聚类原则,最终选择不经任何处理的原光谱参加建模。(2)研究了有效的光谱特征提取方法。用小波分析不仅能提取敏感的光谱特征信息,而且能够有效的降低高维空间数据,与支持向量机结合而成的新方法是一种有效的识别方法。利用db6小波变换得到的77个小波系数作为支持向量机模型的输入。当分解尺度为5时,数据点数由原来的2127个减少至77个(3)研究了支持向量机的模型参数选择。首先用没有内部参数的线性函数作为核函数,改变惩罚因子C,以获得最小的MSE为准则,确定了最佳的误差惩罚因子C的值为1000；然后以不同的核函数以及核函数的内部参数建立支持向量机模型,通过试验发现当核函数为径向基函数,其参数γ为16时,所建立的模型识别率达98.45%。更多还原

【Abstract】 With the development of social economy and the enhancement of living standards, principal contradiction between rice supply and demand was transformed from insufficient quantity to unfavourable quality. Therefore, it was significant to establish rapid rice quality detection method for the purpose of rice breeding, food processing and trade in agriculture.Chalkiness, amylose content and aging of rice have key influence on edible quality of rice. Therefore, in this paper, rapid detection methods for these three characters were studied based on uncertainty artificial intelligence and Chemometrics theory, using digital image processing technology, spectral analysis technology, wavelet analysis and pattern recognition technology.In order to improve identification accuracy and adaptability for chalkiness in computer vision system, the main content about chalkiness rapid detection method were as follows:(1) Image acquisition environment under visible light was established. A computer vision system for image acquisition was set up, analyzing such factors on rice image quality as light source, voltage, and background, and finally the best environmental condition for rice image acquisition was determined. Under transmission light, the best voltage was 6.4V, the best light was LED and the background color was light blue. Under reflected light, the best voltage was 6.0V, the best light is LED and the background color was deep blue.(2) Histogram of the paddy rice image acquired under the transmitted light and the reflected light was analyzed, the appropriate image denoising method and image segmentation algorithm were studied. The result of analyzing histogram of different color space showed that the gray scale distribution of image collected under the reflected light was suitable for chalkiness recognition. Weighted average value filter template was designed in order to not only eliminate the noise but also protect the boundary. The image segmentation algorithm-ostu method was discussed and the rice image rectangular region size was determined based on it.(3) The adaptive chalkiness recognition algorithm was studied. This method was introduced on the basis of the uncertainty artificial intelligence theory and the cloud model. In this method, chalkiness and non-chalkiness were defined as two qualitative concepts, and then they were expressed by an asymmetrical cloud and a symmetrical cloud separately. Chalkiness cloud and non-chalkiness cloud were described by two groups of digital characteristics. Artificial estimation method, the fixed threshold method and the cloud classification method were used in chalkiness area detection under the same voltage value to test the accuracy of cloud classification method. The result showed that the cloud classification’s precision was higher than artificial estimation method, because the average value of deviation between cloud classification and fixed threshold method (i.e. accurate value) was 0.97, while average value of deviation between artificial estimation method and fixed threshold method(i.e. accurate value) was 1.93. Artificial estimation method and cloud classification method were used in chalkiness area detection under the different voltage value to test the adaptability of cloud classification method. The result showed that the cloud classification method’s adaptability was better than the artificial estimation method.In order to establish the near-infrared spectrum quantitative analysis model with good stability and high prediction precision for paddy rice amylose content, the content mainly studied were as follows:(1) The influence of spectrum collecting parameter to the paddy rice near-infrared spectral response characteristic was studied. The spectrums were collected for paddy rice with the same amylose content under the different parameter. The best collecting parameter was determined according to the result of spectrum statistical analysis. The scanning times was 64, the resolution was 8cm"1, and the indoor temperature was 15℃(2) Abnormal spectrum removing method and the spectrum pretreatment method were studied for rice spectrums. In order to optimize the calibration set of sample and enhance the precision of the model,18 abnormal spectrums which were produced from subjective and objective factors were removed from calibration sample set based on mahalanobis distance and forecast concentration residual criterion. The factors such as baseline drift and shift, instrument’s random noise, the stray light had disturbance to the spectrum. In order to eliminate disturbance and improve signal-to-noise ratio, many kinds of methods were used to process the paddy rice spectrums and their influence to model result were compared. The first derivative combined with the SG convolution smoothing method was determined through comparison of model evaluation index.(3)The influence of quantitative analysis method to the model effect was studied. Multiple linear stepwise regression, principal component regression and partial least square regression were used to develop models on spectrums processed by first derivative combined with the SG convolution smoothing method. The model developed by partial least square regression resulted in the best stability, the highest correlation coefficient of the predicted value and the actual value, and the smallest root mean square error of prediction. The correlation coefficient of the predicted value and the actual value was 98.96%, the root mean square error of calibration was 0.62, the root mean square error of prediction was 1.19, and the root mean square error of cross validation was 1.58.In order to study the potential of near-infrared reflectance spectroscopy for detecting aging rice, the main content about aging rice rapid detection method are as follows:(1) The near-infrared spectral response characteristic of the aging rice and the non-aging rice were studied. The influence of different spectrum pretreatment method to the cluster effect was studied using the principal components analysis. The results showed that it was feasible to use paddy rice’s near-infrared spectrums to recognize aging rice qualitatively. Original spectrums were selected to participate in the model according to the cluster principle that the distance in class was the smallest and the distance between classes was the biggest.(2) The effective characteristic extraction method was studied for spectrums of rice. Wavelet analysis could not only withdraw the sensitive signature information of spectrum, but also reduce the dimension of data effectively. Therefore, wavelet analysis combined with support vector machine method provided an effective pattern recognition method.77 wavelet coefficients which obtained from the db6 wavelet transformation were taken as the input of the support vector machine model. When the decomposition size is 5, the number of data points of each spectrum was reduced from 2127 to 77.(3) The parameter choice of support vector machine model was studied. Firstly, linear function was took as the nuclear function without internal parameter, and the model was developed with changing value of penalty factor. The results showed that the best value of error punishment factor C was 1000. Secondly, support vector machine model was developed using different nuclear function as well as the nuclear function’s internal parameter. The experiment results showed that the identification accuracy was 98.45%when the nuclear function is the redial basis function and its parameterγis 16.更多还原