基于多层神经网络的兜兰花分类研究

发布时间：2018-03-05 19:08

本文选题：兜兰花　切入点：颜色矩　出处：《云南大学》2016年博士论文　论文类型：学位论文

【摘要】：花卉分类是图像处理以及自动化分类领域研究的主要研究课题之一。多年来研究人员在这个课题中投入了大量精力,同时也取得不少的研究成果。显然,在未知花卉品种的情况下,只依据花卉图片来识别花卉本身是很有挑战的问题。通常,自然环境中的花卉,在不同的天气和时间,所接收的光照不同,而且同一种花也会有大小和质地之间的变化,这些因素往往会影响花卉分类的结果。如何解决这些问题并最终实现花卉的自动分类,对于那些对花卉感兴趣的用户具有极大的意义,同时也能帮助植物学家和花卉专家高效准确地识别花卉种类。对于花卉的分类问题,论文中选择了一种在泰国很有名的兜兰作为研究对象。兜兰是一种濒危植物,它是兰花的一种,花色丰富,花型奇特,非常有趣的是它一株只开一朵花。但是不同种类的兜兰却有着相似的外观,这使得对兜兰进行分类难度很大。为了实现兜兰的分类,论文利用多层神经网络分类器模型,根据花卉图片的视觉内容进行分类,选取兜兰的颜色特征以及基于分割的分形纹理分析(SFTA)特征作为分类的特征向量。论文的主要贡献如下：(1)建立兜兰花数据库。数据库包含1100幅11种兜兰花图片,与前人的数据集(200幅5种兜兰花)相比,包含的品种更多。此外,兜兰花在中国较为稀少,由于缺少样本,相关的花卉分类研究通常会选择常见的花卉品种,这使得兜兰花的分类研究成了新课题。(2)提出了一种新的兜兰花分类模型一基于多层神经网络的兜兰花分类模型。在分类模型的特征选择上,论文提取了多种特征。在颜色特征上,用到了2种颜色特征：颜色矩和颜色直方图,实验中对比了多种颜色空间如RGB, CIE XYZ, YCbCr以及HSV颜色空间在分类器上的表现,实验结果表明：使用在HSV颜色空间下的颜色特征分类效果最理想；在纹理特征提取上采用基于分割的分形纹理分析(SFTA)算法,该算法与以前的纹理特征提取算法相比,在质量和效率上表现更优。在分类器的选择上,先前的许多研究工作用到的分类器,有随机森林,支持向量机(sVM),人工神经网络(ANN)等。除此之外,论文还对比了一些著名的分类器例如朴素贝叶斯算法、 k近邻分类算法、C4.5决策树算法(J48),序列最小优化SMO算法和多层神经网络(MLP)。实验结果显示,MLP分类器在所有的分类器中表现最为理想,它的平均分类准确率达到了97.64%,所以本论文最终选用MLP分类器。论文的实验结果表明,兜兰花分类模型的准确率令人满意,该模型能帮助植物学家对兜兰进行识别分类,并能为植物学家选种育种提供帮助。在未来,可将该模型应用于花卉图片检索以及对不同类型的花卉分类工作,而且很容易拓展到类似的应用中。
[Abstract]:Flower classification is one of the main research topics in the field of image processing and automatic classification. Over the years, researchers have invested a lot of energy in this subject, and have also achieved a lot of research results. In the case of unknown varieties of flowers, it is a challenging problem to identify flowers only on the basis of pictures of flowers. Usually, flowers in the natural environment receive different light in different weather and time. And the same kind of flower also has the change between the size and the texture, these factors often influence the flower classification result. How to solve these problems and finally realize the automatic classification of flowers, It is of great significance to those who are interested in flowers and can also help botanists and florists identify flower species efficiently and accurately. In this paper, we selected a kind of Daurus, which is very famous in Thailand, as an endangered plant. It is a kind of orchid with rich flowers and peculiar flowers. It is very interesting that it has only one flower per plant. However, the different species have similar appearance, which makes it very difficult to classify the orchid. In order to realize the classification, the paper uses the multi-layer neural network classifier model. According to the visual content of flower pictures, The color features of Cymbidium and the fractal texture analysis (SFTA) feature based on segmentation are selected as the feature vectors of the classification. The main contributions of this paper are as follows: 1: 1) the database is established. The database contains 1100 pictures of 11 species of Cymbidium. It contains more varieties than previous data sets of 200 pieces of five species of cymbidium. In addition, the orchids are rare in China, and because of the lack of samples, related flower classification studies usually select common flower varieties. This makes the research of orchid classification become a new subject. (2) A new classification model of orchid, a classification model based on multi-layer neural network, is proposed. In this paper, two kinds of color features are used: color moment and color histogram. In the experiment, the performance of many color spaces such as RGB, CIE XYZ, YCbCr and HSV color space in classifier is compared. The experimental results show that the color feature classification in HSV color space is the best, and the fractal texture analysis algorithm based on segmentation is used in texture feature extraction, which is compared with the previous texture feature extraction algorithm. Better performance in quality and efficiency. In the selection of classifiers, many previous studies have used classifiers such as random forests, support vector machines (SVMs), artificial neural networks (Ann), etc. The paper also compares some famous classifiers such as naive Bayes algorithm, k-nearest neighbor classification algorithm, C4.5 decision tree algorithm, sequence minimum optimization SMO algorithm and multilayer neural network. The best performance in the class, Its average classification accuracy is 97.64%, so the MLP classifier is used in this paper. The experimental results show that the accuracy of the orchid classification model is satisfactory, and the model can help botanists to recognize and classify the orchid. In the future, the model can be applied to the image retrieval of flowers and the classification of different types of flowers, and can be easily extended to similar applications.
【学位授予单位】：云南大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】