【作者】 贺鹏；

【导师】 黄林；

【作者基本信息】 西北农林科技大学 ， 计算机应用技术， 2008， 硕士

【摘要】 为了有效解决传统植物机器识别对象过宽泛、分类特征较单一且正确识别率较低的问题,提出一种专门针对阔叶树的机器识别方法。该方法将阔叶树叶片的形状特征与其纹理特征相结合,构成一个针对阔叶树叶片图像的综合特征向量,以期更为便捷、快速、高效地对阔叶植物进行计算机自动分类识别。首先提取叶片样本的综合特征信息,然后以概率神经网络(Probabilistic Neural Networks PNN)为分类器对所得到的特征信息进行训练,训练好的网络可以用来识别阔叶的类别,从而确定相应阔叶树的种类。本研究有效提取了含有9个分量的阔叶树叶片综合特征向量,通过对PNN分类器的训练,实现了30种阔叶树的快速机器识别,平均正确识别率达98.3%。比较测试表明:若去掉叶片纹理特征,单以其形状特征作为识别依据,平均识别率仅为93.7%。实验证明针对阔叶树叶片的综合特征识别方法,有效弥补了传统单特征识别方法的不足,使得识别效率得到了较大的提高。本文首次提出基于叶片综合特征的阔叶树机器识别概念,且就此做了比较系统的研究,其中主要的研究工作及结论如下:(1)本文以阔叶树叶片为具体对象,将其形状特征与纹理特征相结合,挖掘并构建出阔叶综合特征信息,并就此综合特征信息的提取做了较深入的研究。文章结合植物学相关理论对叶脉的脉序类型做了一定的研究,且首次将脉序类型的标定值作为叶片的一个纹理特征应用于阔叶树的机器识别中。(2)本文利用已有方法对阔叶树叶片的叶脉图像进行了提取,并在此前提下提出一种改进的差分盒维数算法,且利用此算法对所提取出的叶脉图像进行分形维数的计算,测试结果表明:用此算法求取的叶脉图像分形维数有效地克服了传统分形维数求取方法的不足之处,使得纹理图像的分形维数具有良好的动态范围,结果的准确性得到了较大的提高。(3)首次将概率神经网络(PNN)应用于专门针对阔叶树的机器识别中,且提出一种基于PCA和优化平滑因子的双管道改进PNN分类器模型。(4)应用Microsoft VC++ 6.0和Matlab 6.5,设计开发了一个针对阔叶树的计算机识别系统,该系统可以对阔叶树叶片图像进行全面而合理的预处理,对预处理后的叶片图像可以进行综合特征的准确提取,最后利用所提取到的综合特征信息完成PNN分类器的训练及识别工作,平均正确识别率达98.3%。更多还原

【Abstract】 The study is to effectively solve the problem that the objects of traditional plant identification were too broad and the classification features of it were usually not synthetic and the recognition rate was always slightly low. This paper gives one recognition approach, in which the shape features and the texture features of the leaves of broad-leaved trees combine, composing a synthetic feature vector of broad leaves and hoping to realize the computer automatic classification towards broad-leaved plants more convenient, rapidly and efficient. Firstly, the synthetic features of leaf were extracted; Secondly the values of synthetic features of leaf, which had been extracted, were inputted into a classifier, which is Probabilistic Neural Network(sPNN), to be trained; Finally, the PNN trained well could work to classify the broad leaves and the corresponding broad-leaved trees. The synthetic feature vector of broad leaf, which includes 9 elements, was extracted efficiently. By training the classifier PNN, the rapid recognition for thirty kinds of broad-leaved trees was realized and the average correct recognition rate reached 98.3%. Comparison tests demonstrated that if the shape features of broad leaf solely worked as the recognition features without the texture features, the average correct recognition rate just reached 93.7%. The synthetic feature recognition method against broad-leaved trees has effectively made up the drawbacks of traditional recognition method towards unitary feature, strongly advancing the recognition rate.This paper has demonstrated the notion for the first time that the machine recognition for broad-leaved trees based on leaves’synthetic characteristics, and has made a systematic research about this notion. The following points are the major studys and conclusions of this paper:(1) This paper made the shape feature and the texture feature of the leaves of broad-leaved trees combine, mining and constructing synthetic feature information of broad leaves, and has done a deep research on the extraction of this synthetic feature information. What’s more, based on corresponding botany theory, the neuration types of vein images of broad leaves have been done some research, and, innovatively, the label values of neuration types were applied to the broad-leaved trees recognition system as a texture feature of broad leaves.(2) The vein image of broad leaf has been extracted applying existed method, after which, an improved difference box dimension algorithm were proposed to compute the fractal dimension of vein image. The test result demonstrated that the new algorithm could effectively overcome the drawbacks of the traditional fractal dimension computing method, making the fractal dimension of texture image have a good dynamic range, and the accuracy of results has been better enhanced.(3) Probabilistic Neural Networks (PNN) was first time applied to the broad-leaved trees recognition, what’s more, this paper designed one double-pipes improved PNN classifier model based on PCA and optimization smoothing factor.(4) The broad-leaved trees recognition system has been designed and developed with Microsoft VC++ 6.0 and Matlab 6.5. This system can give comprehensive and reasonable pre-processing to broad leaf images and get accurate extraction of synthetic features of pre-processed leaf image and finally accomplish the training and identification task of PNN classifier by using the information of synthetic features. The average correct recognition rate reached 98.3%.更多还原