【作者】 刘静波；

【导师】 殷涌光；

【作者基本信息】 吉林大学 ， 农业机械化工程， 2004， 博士

【摘要】 人体所需要的许多营养都来自于蔬菜。科学家通过对各类蔬菜营养成分的分析研究，发现蔬菜的营养价值与其颜色有一定关系。掌握和控制蔬菜加工过程中的颜色变化，将颜色进行量化具有十分重要的意义。目前颜色研究中色素成分的分析过程不仅繁琐，而且不能给出人眼实际感觉到的颜色，很难令人满意。而利用计算机视觉系统对产品色泽做出评价，不仅可以克服人眼疲劳造成的差异和评价的局限性，同时还可以利用产品各部分颜色的不同做出相应判断。由于硬件条件的限制，目前国内仍没有商品化的在线检测系统出现，仍采用人工进行颜色分级。为此，本文建立了计算机视觉系统，对物体的颜色、色差进行采集、处理、显示和评价，本课题是军需部项目“舰艇蔬菜保鲜技术”课题的一部分，本论文主要针对以下几个关键技术作了研究：1、在了解人类视觉的构成、信息处理过程的基础上，建立了一套用于颜色量化的计算机视觉系统。测量物体表面颜色，必须在一定光源下进行。不同光源光谱功率分布不同，因此在它们的照射下，物体表面呈现为不同颜色。CIE规定的标准照明体A和D65在CIE1976L*U*V*色空间上颜色宽容量范围接近圆形，因此该色品图上两个颜色点的位置和距离能够较正确地反映出两者的知觉差异，并可用色差这个概念较正确地判断色知觉差异，通过实验发现D65光源更适宜测试绿色蔬菜的颜色变化，因此在本论文中选择D65光源用于颜色量化系统。2、使用1980A彩色亮度计（美国）对颜色量化系统进行标定，彩色亮度计与摄像机对同一标准色卡同步测量，摄像机输出的RGB值，彩色亮度计输出XYZ值，采用1980A彩色亮度计测试标准色卡的测量值作为计算机图像采集系统目标色卡的“真实”颜色值；而摄像机是一个非线性响应，摄像机RGB色空间与CIEXYZ色空间之间的转换关系也是一个复杂的非线性关系，建议利用神经网络完成这一转换。3、在颜色理论的基础上建立了颜色量化软件系统，本系统采用的蔬菜图像文件均为BMP图像文件格式。由于全色谱的样本数量过多，使神经网络训练时间长、收敛速度慢、误差控制困难。因此为了提高预测精度，本文将整个色谱范围内的样本集，分成独立的几个子集，各子集应用各自独立的神经网络进行训练，并给出了绿色神经网络的训练过程。 <WP=130>4、采用三层BP网络实现摄像机输出RGB颜色空间与CIEXYZ色空间转换，其中隐层含有10个节点，传递函数采用tansig函数；输出层传递函数选用purelin函数,得到的训练误差为3.50324×10-5，测试误差为1.4063×10-4，结果满意，可以认为该神经网络可以用来实现这个关系映射。确定了用于RGB和XYZ颜色空间转换的BP网络结构，求出了该神经网络的权值和阈值，在此基础上将神经网络嵌套到Vc编制的程序中，为下一步的颜色量化打下了一个坚实的基础，本文给出了其中的部分代码。5、彩色图像分割可以通过计算图像中色彩信息的统计特性来实现，而无需依赖于待分割区域的几何形状。对于绿色蔬菜来说利用RGB颜色空间中的G分量无法将蔬菜图像和背景分开，而在H分量的直方图上，蔬菜色度H变化范围在60～135之间，背景分布于两侧，根据这一特点分离蔬菜图像与背景，然后根据区域面积大小去除孤立背景点，得到分割后的蔬菜图像。利用本套系统能够进行颜色量化和跟踪测试样品颜色变化的特点，研究了四种塑料包装袋对蔬菜颜色的影响，从中选择出适用于蔬菜包装的材料。6、对碱液处理后的豆角第一天、一周、二周颜色变化进行测试，算出处理前后的色品差，利用二次通用旋转组合设计得出了加碱量与漂烫时间对豆角第一天、一周、二周颜色影响的数学模型分别为：对方程进行显著性检验表明回归方程的置信度为0.95，且拟和得很好，为以后的生产提供一定的理论基础。7、遗传算法以其简单、鲁棒性强、不受搜索限制性条件约束等特点广泛应用于各类优化问题之中。利用遗传算法求解经过碱处理后第一天、一周、二周豆角数学模型的优化问题，取群体规模数为20，杂交概率为0.7，变异概率为0.04，经过100代群体进化，得到的优化解为：第一天加碱量，漂烫时间，目标的函数值为4.564688；一周加碱量，漂烫时间，目标的函数值为0.002896；二周加碱量，漂烫时间，目标的函数值为1.873810。8、由于Arrhenius模型有不少缺点，因此选定多项式模型来拟合试验数据，通过对实验数据的计算分别得到了色差、色品a*随加热时间变化的数学模型，发现一般多项式达到六次时相关系数等于1，说明曲线和试验数据拟合很好，但存在与实际情况不符合的凹点。9、色差可以反映试验样品处理前后的差别，空白样品的色差值随加热时间的延长逐<WP=131>渐增大，说明绿色逐渐退去变成黄褐色。加锌样品的色差均随加热时间的增加先增加后又逐渐减小，也就是说，通过添加锌离子出现了一个颜色变化由小→大→小的过程；而且随着添加锌离子浓度的增加复绿效果更好，在同一加热时间下添加锌离子浓度越大，样品色差越小。10、加锌样品的a*值均随加热时间的增加先增加后又逐渐减小，也就是说，通过添加锌离子出现了一个由褪色到恢复绿色的过程，而且随着添加锌离子浓度的增加复绿效果越好。加锌750ppm的样品随加热时?更多还原

【Abstract】 A lot of nutrition needed by human come from the vegetable. Through analysing the nutrition composition of vegetable, scientists discover that the nutrition value and its color have certain relations. It has very important meaning to grasp and control the color change in course of vegetable processing, and quantify the color. At present the analysis course of pigment composition in color research is not only fussy but also can’t give person actual color which is felt by eye, it is impossible to satisfy people. But the appraisement for product color using computer visual system not only can conquer the difference caused by the exhausted eye and the limitation of appraisement, but also can make judgment correspondingly using the color difference of each part. Because of the restriction of hardware condition, now the commercial online inspecting system still has not appearance in our country, the color classification of manpower is adopted. Therefore, the computer visual system was established in this paper, which can carry through collection, management, display, and appraisement of vegetable color. This study is a part of the Quartermaster ministry project “Warship Vegetable Fresh-Keeping Technology”.Several crucial aspects are discussed as followed in this paper: 1) A suit of computer vision system used in color quantification measurement was established on the basis of knowing the composition of human vision and the course of information processing. Measuring the color of object surface must carry through under certain light source; the spectrum power distribution of different light source is different, therefore, the color presenting on object surface is different under their shining. The color toleration quantum range of standard illumination body A and D65 recommend by CIE is near round, therefore, the position and distance of two color points can reflect accurately consciousness discrepancy between them on this chromaticity figure, and the color consciousness discrepancy can be judged rightly by the conception color difference. Through the experiment we can discover that D65 light is more suited to test the color change of green vegetable, therefore, D65 light is used in the color quantitative system in this paper.2) Using the 1980A color luminance meter to mark the color quantification system, the color luminance meter and CCD camera synchronally measure the same color card, the output of CCD camera is RGB value, and the color luminance meter is XYZ value. The measure value <WP=134>of standard color card with 1980A color luminance meter is regarded as the real color value of target card for the computer image collection system, and that CCD camera is a non-linear respond, the conversion relation between RGB and CIEXYZ color space is also a complex non-linear relation, we suggest to complete this conversion by neural network.3) On the basis of color theory, the software system is established. All the image files in this system are BMP image format. Because the color card quantities of full chromatogram are too many, the training time of neural network is long, convergence speed is slow, and error is difficult to control. Therefore, for raising the forecast precision, this paper has divided all the sample-collections of full chromatogram into several independent subclasses, each subclass applies its own independence neural network to train, and has given the training course of the green neural network.4) The conversion relation between RGB color space of computer visual system and CIEXYZ color space is realized by three-layer BP network, which hide layer has 10 node number and its transfer function adopts tansig function; transfer function of output layer selects purelin function; the training error is 3.50324×10-5, the testing error is 1.4063×10-4, which satisfies the preconcerted demand, so the neural network can be used in realizing this relation mapping. We find the BP network structure used in the conversion between RGB and CIEXYZ color space, gain the weight and threshold value of the neural更多还原