【作者】 刘璎瑛；

【导师】 丁为民；

【作者基本信息】 南京农业大学 ， 农业机械化工程， 2010， 博士

【摘要】 中国是世界上最大的稻米生产国和消费国,却无法跻身于稻米出口大国之列。其原因之一是我国稻米品质检测技术落后,无法保证出口稻米品质,使我国稻米出口缺乏国际竞争力。本文以稻米米粒为研究对象,针对目前稻米米粒加工、分选等过程中存在的实际问题,研究了基于机器视觉技术的稻米外观品质图像检测原理和方法,构建了稻米品质图像静、动态检测系统,给出了适合稻米在线品质评判的图像处理算法。实验分析了稻米内外品质的相关性,验证了外观品质分选对稻米食味品质的影响。根据稻米外观品质特征研究了稻米分选方法,为进一步开发稻米自动化分选系统奠定了基础。本文的主要内容和结论如下：1.对稻米内外品质相关性进行了初步实验研究,分别完成不同品种和同一品种稻米部分品质相关性实验分析。结果表明,江苏产粳稻其稻米食味品质与胶稠度、垩白度、直链淀粉含量和水分相关,稻米的蛋白质含量与稻米品种相关；单一品种稻米的胶稠度和蛋白质含量随米粒垩白的增多而降低,直链淀粉含量随米粒垩白的增多而升高。单一品种相关性实验测定及方差分析表明,单一品种稻米的外观品质中米粒整碎及垩白大小对直链淀粉含量和胶稠度这两种内在品质有显著影响。因此从分选角度看,剔除破碎米和垩白米能够改善稻米的食味品质。2.构建了稻米外观品质机器视觉检测系统。通过颜色校正和几何标定,系统较好实现了稻米多米粒彩色图像的静、动态获取。对稻米米粒外观图像特征进行描述,给出了完整米、垩白米、破碎、黄米和异型米的定义,按照米粒加工分选中常遇到的米样组合,拍摄了全部为完整米的米样图,全部为垩白米的米样图和各种米样混合在一起的混合米样图,以三种米样图为研究对象对多米粒彩色图像处理算法进行了研究。3.提出了一种基于改进最终测量精度法的彩色图像分割效果评判方法。求取米粒分割后去掉背景的边缘轮廓灰度图,以灰度图的灰度均值和方差作为分割评判准则,分别对三种彩色米样图像进行分割颜色通道和分割方法的选择。经实验验证,在I1颜色通道用最大类间方差法进行稻米多米粒图像分割可以取得较好效果。针对目前垩白米分割算法计算量大、自适应性不强等现状,研究了基于且比雪夫逼近的垩白米垩白区域分割算法,对三种米样图进行了垩白区域提取。结果显示,该算法耗时短、鲁棒性强,实现了不同米样图像的垩白区域自动、准确分割。对垩白米正反两面的垩白区域进行分割提取和面积计算,验证了单目视觉在垩白米检测应用中的可行性。4.提出了一种基于霍特林变换的稻米大小、形状特征提取算法。对目前常用的最小外接矩形法进行改进,通过对江苏产籼稻米粒粒型的测定,比较两种算法的准确度和实时性。结果表明,改进最小外接矩形法单粒计算耗时267ms,误差2.24%；基于霍特林变换法单粒计算耗时31ms,误差1.65%。霍特林变换用于稻米大小、形状特征提取实时性好,准确度高。5.选取江苏产的5种粳稻：武香粳14号、淮稻5号、徐稻3号、宁粳1号和徐稻4号,每个品种稻米选择完整米、垩白米、破碎米和异型米各150粒,黄米在5个品种中共选150粒,共计3750粒,拍摄250张静态图像。根据本文图像处理算法提取稻米米粒的9个大小特征参数、10个形状特征参数和31个颜色特征参数,建立了稻米的图像特征数据库。6.研究了基于多结构神经网络的稻米外观品质评判方法。分别对大小形状特征和颜色特征进行主成分分析,根据结果选取面积、粒型、垩白大小和H值作为网络输入的特征参数,经调试,构建了网络结构为5×(4-4-5-1)的多结构神经网络分类器,并与相同网络复杂度的多层BP神经网络进行分类效果比较。结果显示,多结构神经网络分类器对完整米、垩白米、破碎米、黄米和异型米的识别准确率分别为98.3%,92.4%,97.5%,96%,93%,其平均准确率比多层BP神经网络分类器提高6.4个百分点,并且网络训练耗时短。7.分别拍摄0.08m/s、0.12m/s、0.16m/s和0.2m/s四种传送带运行速度下米粒视频图像,研究了基于改进背景差法的运动稻米图像检测方法,完成了对米粒视频图像的背景自动提取、米粒分割、米粒跟踪和特征提取。将不同速度下提取的稻米大小形状特征与静态特征相比较,根据动态偏差和相对误差选取0.12m/s为本文视频图像采集速度。根据提取的特征进行多结构神经网络评判,对完整米、垩白米、破碎米、黄米和异型米的识别准确率分别为95.2%,89.6%,97.3%,90.5%,82.3%。利用Matlab中的Simulink平台对运动稻米图像检测算法进行仿真实现,并给出了算法优化加速的方法。更多还原

【Abstract】 China is the world’s largest rice producer and consumer countries, but not among the major rice exporting countries. One of the main reasons is technological backwardness of China’s rice quality inspection, which can not guarantee the quality of export rice and make China’s exporting rice lack of international competitiveness. In this paper, japonica rice planted in Jiangsu, for example, were researched based on machine vision technology. According to practical problems existed in the process of rice kernels processing and sorting,static and dynamic rice kernels image capturing machine vision system was constructed, image processing algorithms was given to detect rice appearance quality on-line. Experiments were done to analysis relevance of rice’s internal and external quality and verify the appearance quality sorting of rice affected its taste quality. According to characteristics of rice appearance quality, rice kernels grading methods were given, which laid the foundation of developing commercial rice automated grading system. The main contents and conclusions of this paper were as follows:1. Machine vision inspection system of rice appearance quality was constructed to achieve multi-kernel rice color image in the static and dynamic state, through color correction and geometric calibration. The appearance features of rice kernels image was described and the definitions of sound and whole rice, chalky rice, broken rice, yellow rice, and off-type rice were given. According to rice kernels mix mode encountered at processing of rice grading, three kinds of image like all the sound and whole rice image,all the chalky rice image and five classes of rice mixture image were captured to research processing algorithms of multi-kernels color image.2. A method based on improved ultimate measurement accuracy (UMA) was proposed to evaluate color image segmentation performance. Getting the segmentation edge contour gray image removed the background; the gray mean and variance were calculated from this and made as the segmentation judging criterion. Three kinds of color rice sample image were researched on this criterion to select the segmentation method and segmentation color band. The experiments verified that image segmentation using the maximum difference between-cluster with I1 color band can obtain good results. A chalky segmentation algorithm based on Chebyshev approximation was given. Using this method, three kinds of rice samples images were segmented and extracted chalky area. The results showed that this method was time-saving and robust, realized the chalky zone automatic accurate segmentation. The chalky region segmentation and area calculation on rice both sides tested the feasibility of the monocular application.3. An algorithm extracting rice kernels’ size and shape features based on Hotelling transform (HT) was given. The algorithm of minimum enclosing rectangle (MER) widely used at present was improved and compared with HT algorithm at time-consuming and accuracy. Indica rice ratio of length and width planted in Jiangsu province was measured with the above two method. The results showed that the calculation time-consuming percent kernel of improved MER method is 267ms and error is 2.24%, while that of HT method is 31ms and error is 1.65%.The HT method extracted rice kernel image features of size and shape at good real-time and high accuracy.4. Five kinds of Japonica rice planted in Jiangsu province such as Wu Xiang japonica 14, Huai Rice 5, Xu Rice 3, Ning japonica Rice 1 and Xu Rice 4 were selected to research. 3750 rice kernels were selected randomly, which 150 sound and whole rice kernels,150 chalky rice kernels,150 broken rice kernels and 150 off-type rice kernels from every kind of rice,150 yellow rice kernels from all.250 static color images were captured. The algorithms given in this paper were used to extract 9 size features,10 morphological features and 31 color features.5. The method of evaluating rice appearance quality based on multi-structure neural network was researched in this paper. Principal component analysis of size, morphological, and color features gave such four neural network inputs as area, kernel ratio of length to width, chalky area and H value. Primary training of the networks indicated that 5×(4-4-5-1) network was most suitable for the rice grading. The structure of 4-4-5-1 represented those four inputs, four neurons in the first hidden layer, five neurons in the second layer and one in the output layer. The performance of the MSNN classifier was compared against the performance of a multi-layer BP neural network (MLNN) classifier with a similar network complexity. It showed that the accuracy was 98.3% for sound and whole rice,92.4% for chalky rice,97.5% for broken rice,96% for yellow rice and 93%for off-type rice. On the average the MSNN classifier had 6.4% higher recognition accuracy of and shorter training time than the MLNN classifier.6.Rice video images were captured at speed of 0.08m/s,0.12m/s,0.16m/s and 0.2m/s.Rice dynamic image inspection method based on improved background subtraction was researched and it was realized that background automatic extraction, rice kernels segmentation, rice tracking and kernels’ features extraction. Compared features extracted at different speed with the static features,0.12m/s speed was chose for the best suitable speed in terms of dynamic deviation and the relative error. Grading rice using MSNN classifer, the accuracy was 95.2% for sound and whole rice,89.6% for chalky rice,97.3% for broken rice,90.5% for yellow rice and 82.3% for off-type rice. The way of algorithm optimization and acceleration is also given.7. The quality relevance of rice inside and outside was analysized by two group experiments. It showed that Japonica rice taste quality planted in Jiangsu province is related to gel consistency (GC), chalkiness, amylose content (AC) and the moisture content. Protein content (PC) is related to rice length and ratio of length to width. To a single kind of rice, the smaller rice kernels’ chalky area is, the higher GC and PC is; the bigger chalky area is, the higher AC is. The relevant experiment of single kind rice and variance analysis showed that the appearance quality of single kind rice such as length and chalky area has a significant influence on the internal quality such as AC and GC. So removing broken rice and chalky rice by sorting can improve rice taste quality.更多还原