基于图像处理与人工神经网络的烟叶检测系统的研究与应用
|
摘要
烟草行业在我国国民经济中有着重要作用。烟叶分级是烟草行业中的一项基础性工作,按照国家标准对烤烟烟叶的外观质量进行准确的分级,是提高烟草制品质量的关键,对于减少吸烟对身体的危害以及保证农、工、商3个生产环节之间经济利益的合理分配具有重要作用。
目前国内外烟草行业对于烟叶质量的检验与分级都是根据分级标准,依靠人的感官进行经验性判定。近年来,我校的研究团队采用计算机图象处理技术和模式识别技术对烟叶进行特征提取与自动分级,取得了一系列成果。本文在该团队前期研究基础上取得了以下研究成果:
1.基于神经网络的烟叶成熟度、油分检测模型研究烟叶的成熟度和油分是烟叶分级的两个重要指标,无法直接量化提取。本文首先根据烤烟烟叶成熟度和油分与已提取的外观量化因素之间的关系,通过实验选取合适的外观因素组成特征向量,利用反向传播网络和概率神经网络分别建立成熟度和油分两个单指标分类模型,测试分类模型的识别准确率,并对两个网络所建立起来的模型的性能进行比较,确立了概率神经网络在成熟度和油分分类模型中的应用优势。
2.基于不变矩的烟叶形状算法研究本文提取了烤烟烟叶图像的两个不变矩特征,并将不变矩描述作为新增特征分量加入到成熟度和油分两个神经网络分类模型中。通过实验对比,不变矩特征的加入对成熟度分类模型没有起到改善作用,但提高了油分分类模型的识别准确率。
3.基于支持向量机的烟叶生长部位分组研究烤烟烟叶分级的基础是先判断烟叶在烟株上的着生部位,再对不同的部位组进行分级。本文采用支持向量机建立了烤烟烟叶的分组模型,提高了分组准确率。
研究结果表明,采用图像处理、人工神经网络等技术进行烟叶特征提取并建立分组、分级模型具有现实可行性,可以在烟叶质量标准的制定及完善、烟叶质量检验及仲裁以及烟叶分级人员的培训等领域中起到良好的辅助作用。
The tobacco industry plays an important role in China's national economy. Classification of tobacco leaf is a basic work in tobacco industry. It is the key to improve the quality of tobacco products that accurately classifying appearance quality of flue-cured tobacco leaves according to national standards, it also has the important effect in reducing the harm of smoking on the body and ensuring the rational distribution of economic benefits in agricultural, industrial and commercial production links.
The inspection and classification of tobacco leaves in tobacco industry at home and abroad are based on grading standards, determined empirically by relying on people's senses. In recent years, our school's research team used computer image processing and pattern recognition technology for feature extraction and automatic leaf grading and made a series of achievements. On the basis of that team’s preliminary studies this paper made the following findings:
1. Study on tobacco leaf’s maturity, oil detection model based on neural network Tobacco leaf’s maturity and oil which are two important indicators of tobacco leaf classification are unable to be directly quantified and extracted. In this paper, according to the relationship between maturity and oil of flue-cured tobacco leaves with appearance factors already extracted, through the experiment we selected appropriate appearance factors to compose feature vector, established two single - target classification model of maturity and oil by using of back-propagation network and probabilistic neural network, tested the identifying accuracy of classifying models, compared the performance of the model set up by two networks, confirmed the applying superiority of the probabilistic neural network in maturity and oil classifying model.
2. Study on algorithm of tobacco leaf shape based on invariant moments In this paper we extracted two invariant moment features of flue-cured tobacco image, applied invariant moments as new characters in the maturity and oil network classifying model. Through the experimental comparison, the complement of invariant moment features did not serve to improve the maturity model but improve the recognizing accuracy of oil modal.
3. Study on identifying the growing area of tobacco leaves based on support vector machine the base of classifying tobacco leaves is determining the growing area in tobacco plants, then classify the tobacco leaf according to different groups. In this paper, support vector machine had been used to establish the grouping modal of tobacco leaves. The accurate rate of grouping increased.
The results show that using image processing, artificial neural network technology in feature extraction of tobacco leaves and establishing grouping and classifying modal has the feasibility, it is going to play a good supporting role in areas of formulating and improving quality standards of tobacco leaves, verifying and arbitrating quality of tobacco leaves and training classifying staff of tobacco leaves.
目前国内外烟草行业对于烟叶质量的检验与分级都是根据分级标准,依靠人的感官进行经验性判定。近年来,我校的研究团队采用计算机图象处理技术和模式识别技术对烟叶进行特征提取与自动分级,取得了一系列成果。本文在该团队前期研究基础上取得了以下研究成果:
1.基于神经网络的烟叶成熟度、油分检测模型研究烟叶的成熟度和油分是烟叶分级的两个重要指标,无法直接量化提取。本文首先根据烤烟烟叶成熟度和油分与已提取的外观量化因素之间的关系,通过实验选取合适的外观因素组成特征向量,利用反向传播网络和概率神经网络分别建立成熟度和油分两个单指标分类模型,测试分类模型的识别准确率,并对两个网络所建立起来的模型的性能进行比较,确立了概率神经网络在成熟度和油分分类模型中的应用优势。
2.基于不变矩的烟叶形状算法研究本文提取了烤烟烟叶图像的两个不变矩特征,并将不变矩描述作为新增特征分量加入到成熟度和油分两个神经网络分类模型中。通过实验对比,不变矩特征的加入对成熟度分类模型没有起到改善作用,但提高了油分分类模型的识别准确率。
3.基于支持向量机的烟叶生长部位分组研究烤烟烟叶分级的基础是先判断烟叶在烟株上的着生部位,再对不同的部位组进行分级。本文采用支持向量机建立了烤烟烟叶的分组模型,提高了分组准确率。
研究结果表明,采用图像处理、人工神经网络等技术进行烟叶特征提取并建立分组、分级模型具有现实可行性,可以在烟叶质量标准的制定及完善、烟叶质量检验及仲裁以及烟叶分级人员的培训等领域中起到良好的辅助作用。
The tobacco industry plays an important role in China's national economy. Classification of tobacco leaf is a basic work in tobacco industry. It is the key to improve the quality of tobacco products that accurately classifying appearance quality of flue-cured tobacco leaves according to national standards, it also has the important effect in reducing the harm of smoking on the body and ensuring the rational distribution of economic benefits in agricultural, industrial and commercial production links.
The inspection and classification of tobacco leaves in tobacco industry at home and abroad are based on grading standards, determined empirically by relying on people's senses. In recent years, our school's research team used computer image processing and pattern recognition technology for feature extraction and automatic leaf grading and made a series of achievements. On the basis of that team’s preliminary studies this paper made the following findings:
1. Study on tobacco leaf’s maturity, oil detection model based on neural network Tobacco leaf’s maturity and oil which are two important indicators of tobacco leaf classification are unable to be directly quantified and extracted. In this paper, according to the relationship between maturity and oil of flue-cured tobacco leaves with appearance factors already extracted, through the experiment we selected appropriate appearance factors to compose feature vector, established two single - target classification model of maturity and oil by using of back-propagation network and probabilistic neural network, tested the identifying accuracy of classifying models, compared the performance of the model set up by two networks, confirmed the applying superiority of the probabilistic neural network in maturity and oil classifying model.
2. Study on algorithm of tobacco leaf shape based on invariant moments In this paper we extracted two invariant moment features of flue-cured tobacco image, applied invariant moments as new characters in the maturity and oil network classifying model. Through the experimental comparison, the complement of invariant moment features did not serve to improve the maturity model but improve the recognizing accuracy of oil modal.
3. Study on identifying the growing area of tobacco leaves based on support vector machine the base of classifying tobacco leaves is determining the growing area in tobacco plants, then classify the tobacco leaf according to different groups. In this paper, support vector machine had been used to establish the grouping modal of tobacco leaves. The accurate rate of grouping increased.
The results show that using image processing, artificial neural network technology in feature extraction of tobacco leaves and establishing grouping and classifying modal has the feasibility, it is going to play a good supporting role in areas of formulating and improving quality standards of tobacco leaves, verifying and arbitrating quality of tobacco leaves and training classifying staff of tobacco leaves.
引文
[1] Splinter W E Electronic micrometer continuously monitors plant stem diameter[J]. Agricultural Engineering, 1969, 50:220-221.
[2]孙红,孙明,王一鸣.植物生长机器视觉无损测量研究综述[J].农业机械学报, 2006, 37(10): 181-185.
[3]何东健,张海亮,宁继锋等.农业自动化领域中计算机视觉技术的应用[J].农业工程学报, 2002, 18(2): 171~175.
[4]王传宇,赵明,阎建河.基于双目立体视觉的苗期玉米株形测量[J].农业机械学报, 2009, 40(5): 145-148, 144.
[5] Thomas C E. Techniques of image analysis applied to the measurement of tobacco and related products[R]. 42nd Tobacco Chemists’Research Conference, 1988.
[6] J K M MacCormac, On-Line Image Processing for Tobacco Grading in Zimbabwe[J], IEEE, 1993: 327-331.
[7] CHO H K ,PAEK K H, Feasibility of grading dried Burley tobacco leaves using machine vision, Journal of the Korean society foe agricultural machinery[J], 1997, 22(1): 30-40.
[8]张建平,数字图像处理在农产品质量检测中的应用-烟叶质量的自动分级[D].镇江:江苏理工大学, 1994.
[9]张建平,吴守一,方如明.农产品质量的计算机辅助检验与分级(第Ⅰ报)——烟叶外观品质特征的定量检验[J],农业工程学报, 1996, 12(3): 158-162.
[10]张建平,吴守一,方如明等.农产品质量的计算机辅助检验与分级(第Ⅱ报)——烟叶自动分级模型的建立与训练[J].农业工程学报, 1997, 13(4): 179-183.
[11]韩力群,何为,苏维军等.光机电一体化烤烟烟叶图像采集系统的研发[J].微计算机信息, 2006, 22(3): 228-229, 200.
[12]闫瑞琼,韩力群,陈晋东.计算机技术在烟叶检测与分级领域的应用[J].烟草科技, 2001, 3: 13-15.
[13]韩力群,何为,段振刚,等.烤烟烟叶自动分级的智能技术[J].农业工程学报, 2002, 18(6): 173-175.
[14]韩力群,周文.基于图像处理技术的烤烟烟叶质量检测研究.自动化学报, 2000增刊B: 178-181.
[15]韩力群,苏维军,段振刚等.基于拟脑智能系统的考验烟叶分级研究[J].农业工程学报, 2008, 24(7): 137-140.
[16]周文,韩力群,李锐.计算机图象处理技术在烤烟烟叶形状特征提取中的应用[J].烟草科技, 2000, 1: 12-13.
[17]郑建冬,伍铁军,商庆清.基于图像处理的烟叶叶片结构分类算法[J].东北大学学报, 2007, 33(3): 282-286.
[18]郑志军,张乐年,李世平等.基于图像处理的烟叶分拣系统的设计与实现[J].计算机测量与控制, 2006, 14(4): 467-469.
[19]李强,杨晓京,魏岚等.基于计算机视觉的烟叶分离系统[J].现代制造工程, 2006, 5: 101-103.
[20]贺英,冯天瑾,曹均阔. Kohonen网络在烟叶动态分类中的应用[J].中国海洋大学学报, 2004, 34(1): 121-127.
[21]彭晏飞,刘树启.基于BP算法的纹理图像合成研究[J].世界科技研究与发展, 2009, 31(3):462-464.
[22]张震南.人工神经网络技术在语音识别中的应用[J].甘肃科技纵横, 2008, 37(4): 21, 76.
[23]孙宁,孙劲光,孙宇.基于神经网络的语音识别技术研究[J].计算机与数字工程, 2006, 34(3): 58-60.
[24]潘旦光,黄慧光,宋俊磊.基于失效树理论的结构多位置损伤检测[J].金属世界, 2009(z1):34-38.
[25] C B Yun, E Y Bahng. Sub-structural identification using neural networks[M]. 2000.
[26] Y Ikeda. Y Saito. Shape judgment of agricultural products via neural network judgment of shape and size of pimento. Journal of the Japanese Society of Agricultural Machinery[J], 1993, 55(3): 59-64.
[27]李洪文,李慧,何进等.基于人工神经网络的田间秸秆覆盖率检测系统[J].农业机械学报, 2009, 40(6): 58-62.
[28]彭黔荣,蔡元青,王东山等.根据常规化学指标识别烟叶品质的BP神经网络模型[J].中国烟草学报, 2005, 11(5):19-25.
[29]王强,陈英武,李孟军.卷烟焦油量的支持向量机预测[J].烟草科技, 2007, 10: 5-8.
[30]焦艳华,张雪萍,林楠.支持向量机在烤烟烟叶自动分级中的应用[J].微计算机信息, 2009, 25(8-1): 195-196, 167.
[31]中国烟叶生产购销公司.烤烟分级国家标准培训教材[M].北京:中国标准出版社, 2004.
[32] Specht D F. Probabilistic neural networks for classification mapping, or associative memory[C]. Proceedings of IEEE International Conference on Neural Networks, 1988, 1:525-532.
[33]雷可君.基于小波包变换与PNN神经网络相结合的指纹识别系统[D].湖南:湖南大学, 2007.
[34]高全华,王晋国,孙锋利.基于Pseudo-Zernike不变矩的PNN车牌汉字识别[J].计算机工程, 2009, 35(4): 196-198.
[35]杨鼎强,肖淑萍,蒋加伏.基于差异演化概率神经网络的纹理图像识别[J].计算机工程与应用, 2008, 44(11): 179-181, 198.
[36]黄林,贺鹏,王经民.基于概率神经网络和分形的植物叶片机器识别研究[J].西北农林科技大学学报, 2008, 36(9): 212-217.
[37]宋薇,陶智,顾济华等.基于概率神经网络的汉语耳语音识别系统[J].计算机工程与应用, 2008, 44(17): 148-150.
[38]史忠植.神经网络(M).北京:高等教育出版社, 2009: 203-211.
[39] Vladimir N.Vapnik,统计学习理论的本质[M]. (张学工,译).北京:清华大学出版社, 2000.
[40] HWANGKJ.KIM CW,KIM CH Studies on the change of chemical components of flue-cured tobacco with maturity[C]. Coresta Congress,Agro-Photo Groups, 1981:10.
[41]陆天胜,刘汉传,邵伏文等.烤烟40级不同成熟度与内外观质量研究[J].安徽农业科学, 1994, 22(3): 256-258.
[42]王军,高远峰,王闯.叶成熟度与烟叶品质的关系探讨[J].河北农业科学, 2008(6):16-18.
[43]李章海,刘登乾,韩忠明等.烤烟油分与烟叶理化特性关系的初步研究[J].安徽农业科学, 2008, 36(3): 1088-1089, 1105.
[2]孙红,孙明,王一鸣.植物生长机器视觉无损测量研究综述[J].农业机械学报, 2006, 37(10): 181-185.
[3]何东健,张海亮,宁继锋等.农业自动化领域中计算机视觉技术的应用[J].农业工程学报, 2002, 18(2): 171~175.
[4]王传宇,赵明,阎建河.基于双目立体视觉的苗期玉米株形测量[J].农业机械学报, 2009, 40(5): 145-148, 144.
[5] Thomas C E. Techniques of image analysis applied to the measurement of tobacco and related products[R]. 42nd Tobacco Chemists’Research Conference, 1988.
[6] J K M MacCormac, On-Line Image Processing for Tobacco Grading in Zimbabwe[J], IEEE, 1993: 327-331.
[7] CHO H K ,PAEK K H, Feasibility of grading dried Burley tobacco leaves using machine vision, Journal of the Korean society foe agricultural machinery[J], 1997, 22(1): 30-40.
[8]张建平,数字图像处理在农产品质量检测中的应用-烟叶质量的自动分级[D].镇江:江苏理工大学, 1994.
[9]张建平,吴守一,方如明.农产品质量的计算机辅助检验与分级(第Ⅰ报)——烟叶外观品质特征的定量检验[J],农业工程学报, 1996, 12(3): 158-162.
[10]张建平,吴守一,方如明等.农产品质量的计算机辅助检验与分级(第Ⅱ报)——烟叶自动分级模型的建立与训练[J].农业工程学报, 1997, 13(4): 179-183.
[11]韩力群,何为,苏维军等.光机电一体化烤烟烟叶图像采集系统的研发[J].微计算机信息, 2006, 22(3): 228-229, 200.
[12]闫瑞琼,韩力群,陈晋东.计算机技术在烟叶检测与分级领域的应用[J].烟草科技, 2001, 3: 13-15.
[13]韩力群,何为,段振刚,等.烤烟烟叶自动分级的智能技术[J].农业工程学报, 2002, 18(6): 173-175.
[14]韩力群,周文.基于图像处理技术的烤烟烟叶质量检测研究.自动化学报, 2000增刊B: 178-181.
[15]韩力群,苏维军,段振刚等.基于拟脑智能系统的考验烟叶分级研究[J].农业工程学报, 2008, 24(7): 137-140.
[16]周文,韩力群,李锐.计算机图象处理技术在烤烟烟叶形状特征提取中的应用[J].烟草科技, 2000, 1: 12-13.
[17]郑建冬,伍铁军,商庆清.基于图像处理的烟叶叶片结构分类算法[J].东北大学学报, 2007, 33(3): 282-286.
[18]郑志军,张乐年,李世平等.基于图像处理的烟叶分拣系统的设计与实现[J].计算机测量与控制, 2006, 14(4): 467-469.
[19]李强,杨晓京,魏岚等.基于计算机视觉的烟叶分离系统[J].现代制造工程, 2006, 5: 101-103.
[20]贺英,冯天瑾,曹均阔. Kohonen网络在烟叶动态分类中的应用[J].中国海洋大学学报, 2004, 34(1): 121-127.
[21]彭晏飞,刘树启.基于BP算法的纹理图像合成研究[J].世界科技研究与发展, 2009, 31(3):462-464.
[22]张震南.人工神经网络技术在语音识别中的应用[J].甘肃科技纵横, 2008, 37(4): 21, 76.
[23]孙宁,孙劲光,孙宇.基于神经网络的语音识别技术研究[J].计算机与数字工程, 2006, 34(3): 58-60.
[24]潘旦光,黄慧光,宋俊磊.基于失效树理论的结构多位置损伤检测[J].金属世界, 2009(z1):34-38.
[25] C B Yun, E Y Bahng. Sub-structural identification using neural networks[M]. 2000.
[26] Y Ikeda. Y Saito. Shape judgment of agricultural products via neural network judgment of shape and size of pimento. Journal of the Japanese Society of Agricultural Machinery[J], 1993, 55(3): 59-64.
[27]李洪文,李慧,何进等.基于人工神经网络的田间秸秆覆盖率检测系统[J].农业机械学报, 2009, 40(6): 58-62.
[28]彭黔荣,蔡元青,王东山等.根据常规化学指标识别烟叶品质的BP神经网络模型[J].中国烟草学报, 2005, 11(5):19-25.
[29]王强,陈英武,李孟军.卷烟焦油量的支持向量机预测[J].烟草科技, 2007, 10: 5-8.
[30]焦艳华,张雪萍,林楠.支持向量机在烤烟烟叶自动分级中的应用[J].微计算机信息, 2009, 25(8-1): 195-196, 167.
[31]中国烟叶生产购销公司.烤烟分级国家标准培训教材[M].北京:中国标准出版社, 2004.
[32] Specht D F. Probabilistic neural networks for classification mapping, or associative memory[C]. Proceedings of IEEE International Conference on Neural Networks, 1988, 1:525-532.
[33]雷可君.基于小波包变换与PNN神经网络相结合的指纹识别系统[D].湖南:湖南大学, 2007.
[34]高全华,王晋国,孙锋利.基于Pseudo-Zernike不变矩的PNN车牌汉字识别[J].计算机工程, 2009, 35(4): 196-198.
[35]杨鼎强,肖淑萍,蒋加伏.基于差异演化概率神经网络的纹理图像识别[J].计算机工程与应用, 2008, 44(11): 179-181, 198.
[36]黄林,贺鹏,王经民.基于概率神经网络和分形的植物叶片机器识别研究[J].西北农林科技大学学报, 2008, 36(9): 212-217.
[37]宋薇,陶智,顾济华等.基于概率神经网络的汉语耳语音识别系统[J].计算机工程与应用, 2008, 44(17): 148-150.
[38]史忠植.神经网络(M).北京:高等教育出版社, 2009: 203-211.
[39] Vladimir N.Vapnik,统计学习理论的本质[M]. (张学工,译).北京:清华大学出版社, 2000.
[40] HWANGKJ.KIM CW,KIM CH Studies on the change of chemical components of flue-cured tobacco with maturity[C]. Coresta Congress,Agro-Photo Groups, 1981:10.
[41]陆天胜,刘汉传,邵伏文等.烤烟40级不同成熟度与内外观质量研究[J].安徽农业科学, 1994, 22(3): 256-258.
[42]王军,高远峰,王闯.叶成熟度与烟叶品质的关系探讨[J].河北农业科学, 2008(6):16-18.
[43]李章海,刘登乾,韩忠明等.烤烟油分与烟叶理化特性关系的初步研究[J].安徽农业科学, 2008, 36(3): 1088-1089, 1105.