【作者】 甘露萍；

【导师】 谢守勇；

【作者基本信息】 西南大学 ， 农业机械化工程， 2009， 硕士

【摘要】 论文利用现代计算机视觉技术,结合地方经济建设,在寻求快速、简便、准确判断鲜烟叶含水量的方法上作了有益的探索。烤烟工艺中,鲜烟叶水分含量与烤烟的烘烤特性有密切关系,新鲜烟叶即湿烟叶含水量准确判断是决策最优烘烤模式和制定合理烟叶烘烤工艺和程序的重要依据。如果将串挂好的不同含水量的湿烟叶合理布置在烤房不同温湿度点热流体流场位置进行统筹烘烤,与随意串挂相比,大大减少烘烤时间,而且克服了烘烤过程中含水量多难烤,含水量少易烤性差的问题。此外提高了烘烤效率、烤烟品质和优质烟叶百分比,使烟叶平均价值增加,煤和其它能源的浪费减少,具有显著的经济效益。如果采用人工检测,劳动强度大、主观性大、干扰大、工作效率低、分级标准不易掌握、分级精度也很难稳定。即使很有经验的受过培训的技师或烟农由于受人主观因素的影响太多,譬如个人的经验,且许多指标呈模糊状态,不能准确判定烟叶水分含量。本文用机器视觉技术将在烟田采集的160幅湿烟叶叶片样本中与含水量相关性大的表象特征检测提取出来,实现特征提取的无损化和处理的快速化,用Elman神经网络科学建立具有一定精度的烟叶含水量非线性评判体系模型,通过模型可以预测叶片实际含水量。为了提高系统精度,进行了系统构建、图像预处理、图像特征的提取、算法的选择等关键技术的理论研究。用Matlab图像处理软件将目标与背景分割,通过比较,选用“无振铃”现象、特性均衡的Butterworth低通滤波器对叶片进行滤波。运用Matlab语言设计颜色、纹理、外形算法进行图像特征提取,分别提取了鲜烟叶的颜色、纹理、外形特征值作为烟叶含水量综合评价指标,以减少单参数对其含水量判定缺陷。用三标度AHP层次分析法对指标综合排序,分析结果表明:叶片均值、宽和面积在整个评价体系中占较大权重,可用作快速判断含水量的指标,将此三个指标作为Elman神经网络的输入,以含水量作为输出目标建立三层神经网络,采用附加动量因子的梯度下降权值/阈值学习函数learngdm对网络进行训练,训练后的网络对测试样本进行含水量预测,测试样本与实际值的相对误差控制在10%范围内,预测精度大于90%,预测结果达到预期目标。利用Matlab中的GUI(Graphic User Interface)设计图像处理仿真平台,制作一个供反复使用且操作简单的专用工具。提供图像输入、图像增强、图像分析及输出鲜烟叶含水量的功能。能根据新鲜烟叶叶片含水量模型预测输入鲜烟叶图片的含水量。研究表明,基于机器视觉的鲜烟叶含水量信息外观无损检测机理可行,检测效果较好,为鲜烟叶烘烤控制工艺和程序的制定提供了依据,为在后期研究中匹配不同烘干曲线从而选择较佳烘干模式中提供了重要依据。还可将检测出的叶片含水量作为田间精准灌溉的依据。此外还可将图像分析出的叶片相关特征数据与叶片营养状况信息指标,譬如氮含量、磷含量进行相关性分析找出简便且准确分析叶片营养状态信息的方法。更多还原

【Abstract】 To judge the fresh tobacco leaf’s water contents quickly, simply, conveniently and accurately, this paper does some meaning exploration, using modern computer vision technology and combining with local economic construction. In the process of flue-cured tobacoo, fresh tobacco leaf’s water contents have near relationships with tobacco leaf’s flue-cured characteristic. It is the important foundation to choose the optimization flue-cured mode and establish the proper technics and flue-cured procedure. Compared with hang discretionarily, if the wet leaf which have different water contents is placed in hot liquid which have different temperature in order to hang orderly ,it does not only the time will be shorter but also the problem that is hard to flue-cured when the leaf contain much water and the character of flue-cured is bad when the leaf contain less water will be solved. In addition, the efficiency and flue-cured quality and the proportion of high quality tobacco will be improved, increasing the mean value of tobacco, decreasing the wasting of coal and other energy sources, which have dramatically economical value.If detected by manpower, the intensity of work is dense, the subjectivity is big and can be easily disturbed.The work efficiency is low, the classification standard is hard to master and also the classification precision is hard to be stabilized. The technician and tobacco farmer was easily influenced by subjective (personal experience) factor even if they have abundance experience and were well trained, and many index take on illegible state, which can not be used to judge the leaf water contents exactly. This paper extract many characters non-destructive and quickly involved with water contents from 160 wet tobacco leaf sample gathered in land, set up a non-linear judge model about leaf water contents using Elman neural network . By which, the leaf ’s real water contents can be detected.To improve the precision of the system, engageing in theory study about key technology with system construction, image preprocessing,image feature extraction, algorithm choice and so on,engaging in image segmentation between the image and the background by MATLAB.The leaf have been filtered by Butterworth filter which does not have "flap bell" phenomenon and have uniform characteristic, design the algorithm about color、vein and outline to draw image characters by MATLAB, extract fresh tobacco leaf’s color、vein and outline as the synthetical judgment indexes, in this way ,to reduce single index’s limitation. Rank the indexes synthetically by the three-demarcation AHP method, The analytical conclusion indicate that the leaf’s width、acreage and mean value take a biggish ratio in the whole judgment system, can be used as the swift judgment indexes of water contents,seting up three layers neural network, taking the three indexes as the input, the water contents as the output. Train the network by learngdm. Predict the water contents of testing sample using the trained network. The relative error between the test sample value and the real value has been controlled below 10%, the predicting precision exceed 90%, the predicting results achieve anticipating goals.Design a professional instrument which can be used repeatedly and operated simply with the image processing emulational platform of the matlab’s GUI, provide function in image input\image enhancement\image analysis and outputting water contents of fresh tobacco leaf, can predict the inputting fresh tobacco leaf’s water contents according to the above model.The research indicate that this nondestructive detecting technique is feasible, the detecting result is accurate, which is the nicer foundation to establish the tectonics and procedure of flue-cured tobacco leaf and to choose the best flue-cured model that fit with different flue-cured curve in later research. In additional, it can be used as the foundation of field irrigation and find out the method about the leaf’s nutritional information simply and exactly, according to the interrelated characteristic data about the leaf and the leaf’s nutritional indexes, such as the contents of nitrogen\ phosphor.更多还原