Quality evaluation of chrysanthemum cut flower using image processing techniques and Kalman neuro

菊花切花的分级或质量评估传统上是由受过分级培训的专家根据他们熟练的敏感性进行的。本研究试图通过对菊花切花的质量进行评分，从传统的评价过程中得出一些定量的标准。结果显示，由两位专家分别给出的两组分数之间存在显著差异。第一次评价与第二次评价也存在较大差异。通过测量切花长度、花到最上节的长度、主茎直径、主茎曲率、平均节间长度、叶和茎的面积、叶的大小，探讨切花的物理特征与专家判定标准的关系。结果表明，大部分被测的身体特征与专家的决策标准有关。这些物理特征似乎不可能直接应用于分级参数，因为没有足够的统计实质性。一定有一些复杂的物理特征之间的组合，使专家判断单个切花的质量。神经网络的一个众所周知的功能是分类器，它可以处理这类问题。在此基础上，选择若干特征作为神经网络的输入参数，神经网络的输出参数为人的评价分数。采用KNT（卡尔曼神经训练）方法对神经网络进行训练。从结果中可以看出，输出值与人类评价分数基本一致。该错误小于由人为双重检查程序导致的人为错误。同时也证实了神经网络结合几个合适的特征进行评价是有效的。

The grading or quality evaluation of chrysanthemum cut flowers is traditionally performed by experts trained in the grading based on their skilled sensibility. In this study, an attempt was made to draw some quantitative criteria from the traditional evaluation process by scoring the quality of chrysanthemum cut flowers. The results revealed that there were significant differences between two sets of scores given by two experts respectively. There were also large difference between the first evaluation and the second one. The measurements were taken for cut flower length, length between flower and the uppermost node, main stem diameter, curvature of main stem, average internode length, area of leaves and stems, and sizes of leaves in order to investigate the relationship between physical features of cut flowers and experts' decision criteria. It seemed that the most of measured physical features were related to experts' decision criteria. No straight applications of these physical features to the grading parameters seem to be possible because there was not enough statistical substantiality. There must be some complex combinations between physical features that make experts decide the quality of individual cut flower. One of the well known functions of neural network is a classifier that can handle this type of problem. Base on the results, several features were selected for input parameters of neural networks whose output parameter was a human evaluation score. The neural networks were trained by KNT (Kalman Neuro Training) method. From the results, it was observed that output value satisfactorily agreed the human evaluation score. The error was less than the human error resulted from the human double check procedure. It was also confirmed that the evaluation by the neural networks with several appropriate features was effective.