基于半监督主动学习的菊花表型分类研究

Chrysanthemum Phenotypic Classification Based on Semi-supervised Active Learning

鉴于人工和专家分类模式的局限性,基于表型的菊花分类存在效率低下的问题。本文采用基于半监督主动学习技术,在已分类菊花数据的基础上,利用未标号菊花样本数据提供的信息,建立了菊花表型分类模型,提升了分类质量和效率。该模型可以不依赖外界交互,利用未标号样本来自动提升菊花分类的质量。为了训练学习模型,本文收集了菊花的表型特征数据,标注了菊花表型类别,并研究了菊花分类属性特征的编码技术。在此数据集上,采用基于图标号传播的半监督学习技术对未标号的菊花数据进行建模,为了提升半监督分类的有效性,在标号传播的基础上使用主动学习技术,采用熵最大策略来选择难以识别的样本,以改进分类质量。在该数据集上进行了试验验证,并进行了试验对比和分析,试验结果表明,本文方法能够较好地利用未标号菊花样本提升分类的精度,随着标号百分比从6.25%升至23%,识别精度达到0.7以上,标号百分比在81.25%时,平均识别精度和召回率分别达到0.91和0.88。

Phenotype-based classification plays an essential role in plant research. Chrysanthemum flower has great momentous economic value and medicinal value,and has feature of morphological and genetic diversity as well. Due to the limitations of the artificial classification model by expert and the characteristic of genetic diversity,phenotype-based classification has been facing great challenges for its research. At present,the technologies and applications of machine learning and artificial intelligence are developing rapidly. With the vehicle of machine learning,the semi-supervised learning technology was employed to provide an effective way for improving the classification performance. This method was based on label propagation of graph model as well as active learning technique. According to this method,a small number of classified chrysanthemum data as well as a large amount of unlabeled chrysanthemum samples were exploited to improve the classification accuracy. This method can automatically make use of the unlabeled samples to improve the quality of chrysanthemum classification without relying on external interactions. The chrysanthemum phenotypic data was collected to train the learning model,and manually annotate the chrysanthemum category information. For exploiting the categorical attribute,the coding skill was studied as well. The label propagation of graph model was utilized by the semi-supervised learning skill for the unlabeled chrysanthemums. In order to improve the effectiveness of semi-supervised classification,active learning technique was applied,which was based on the entropy maximization strategy to select difficult-to-identify samples to improve classification performance further. Extensiveexperiments were conducted and comparisons were made. The experimental results showed that the unlabeled chrysanthemum samples can improve the classification accuracy remarkably,with the labeled ratio increasing from 6. 25% to 23%, the recognition accuracy rapidly reached 0. 7, the average recognition accuracy and recall rate can reach 0. 91 and 0. 88,respectively,when the labeled ratio was81. 25%. In conclusion,semi-supervised based learning for the intelligent identification and effective management of chrysanthemum flowers had great significance in theory and application for the studying of chrysanthemum phenotype.