Prediction of tobacco leaf’s water contents during bulk curing process based on image processing technique | 基于图像处理的烘烤过程中烟叶含水量检测

صينى. 量化烘烤过程中烟叶形态变化的数值特征指标，实现烘烤过程烟叶水分含量的无损检测。以密集烤房中不同烘烤阶段的烟叶为研究对象，先利用图像处理技术提取鲜烟叶及烘烤过程中烟叶图像的颜色特征（红分量（R）、绿分量（G）、蓝分量(B)）及纹理特征(纹理能量、纹理熵、纹理惯性、相关度)，以其为输入指标，分别建立烘烤过程中烟叶含水量的BP神经网络模型和基于遗传算法的最小二乘支持向量机预测模型。用建立的2个模型对烘烤过程中烟叶含水量进行预测，并比较其预测精度。烟叶图像颜色特征R、G、B分量表现出变黄期剧烈上升，定色前期缓慢上升并达到最大值，定色后期至烘烤结束逐渐下降的变化趋势；纹理能量和相关度呈现出变黄前期减小，变黄后期增大，定色及干筋期逐渐减小的趋势；纹理熵、纹理惯性表现出变黄前期增大，变黄后期减小，定色及干筋期逐渐增大的趋势。以烟叶颜色和纹理特征值作为输入变量，建立了烘烤过程中烟叶含水量的BP神经网络预测模型和基于遗传算法的最小二乘支持向量机预测模型，其预测平均绝对误差分别为0.037 4和0.017 0，预测误差标准差分别为0.048 5和0.020 0，前者预测精度略低于后者，但2个模型均可以满足烘烤过程中烟叶水分含量实时检测的需要。图像处理技术可以精确量化烘烤过程中烟叶的形态特征变化；利用建立的BP神经网络模型和基于遗传算法的最小二乘支持向量机模型可以实现对烟叶含水量的精确估测。

إنجليزي. The study was conducted to perform a quantitative research on the change of external morphologic features of tobacco leaves during curing, then realize the nondestructive testing of tobacco leaf’s water contents. In this experiment, tobacco leaves at different stages during bulk curing process and fresh leaves were used as experimental targets. Image processing technology was adopted to extract color characteristics parameters R, G, B and veins characteristics parameters energy, entropy, moment, correlation, then the color characteristics and veins characteristics were taken as the input indexes while BP neural networks, genetic algorithm and least squares support vector machines were used to predict the water contents during bulk curing process. The color characteristics parameters R, G, B obviously increased in the period of yellowing, slowly increased and reached the maximum in the earlier stage of fixing-color, and then decreased from later stage of fixing-color to the end of flue-curing. The energy and correlation decreased in the earlier stage of yellowing, increased in the later stage of yellowing, and then decreased from fixing-color stage to the end of flue-curing. The entropy and moment increased in the earlier stage of yellowing, decreased in the later stage, and then increased from fixing-color stage to the end of flue-curing. Then the color characteristics parameters and the veins characteristics parameters were used as the input to the BP neural network, genetic algorithm and least squares support vector machines with the output of water content. The mean absolute error and root mean square deviation between the measured data and predicted data from BP neural networks were 0.037 4, 0.048 5 respectively, and 0.017 0, 0.020 0 from LS-SVM. The change of external morphologic features of tobacco leaves during curing can be precisely quantitative based on image processing technique, both methods based on BP neural network and genetic algorithm and least squares support vector machines can estimate the water content accurately.