In crop image recognition,when faced with a large quantity of unlabeled data,the traditional manual labeling method requires a large amount of human and material resources.To solve this problem,this study proposes an ...In crop image recognition,when faced with a large quantity of unlabeled data,the traditional manual labeling method requires a large amount of human and material resources.To solve this problem,this study proposes an image recognition method based on a pseudolabeling technique.First,the data are divided into labeled and unlabeled data.The initial network model is trained on labeled data.Then,pseudolabeling of the unlabeled data is predicted,and only the data that satisfy the confidence threshold are regarded as valid pseudolabeling.To convert the unlabeled data into supervised training data,the two types of data are mixed.The training is terminated when the number of remaining unlabeled data satisfies the end condition and when the fivefold crossvalidation method is used to evaluate model performance.Compared with the traditional semisupervised method,the experimental method is simpler and more applicable.Experiments were conducted on rice growth stage recognition and crop weed seedling recognition tasks.The results showed that the proposed method achieved 99.17%accuracy in rice growth stage recognition and a high AUC value of 99.93%in crop weed seedling recognition,which demonstrated excellent performance.Compared with the traditional model,this method not only improves in accuracy but also has better stability and wider applicability and is expected to provide an efficient,accurate and scalable solution for crop image recognition.展开更多
目的随着实际应用场景中海量数据采集技术的发展和数据标注成本的不断增加,自监督学习成为海量数据分析的一个重要策略。然而,如何从海量数据中抽取有用的监督信息,并该监督信息下开展有效的学习仍然是制约该方向发展的研究难点。为此,...目的随着实际应用场景中海量数据采集技术的发展和数据标注成本的不断增加,自监督学习成为海量数据分析的一个重要策略。然而,如何从海量数据中抽取有用的监督信息,并该监督信息下开展有效的学习仍然是制约该方向发展的研究难点。为此,提出了一个基于共识图学习的自监督集成聚类框架。方法框架主要包括3个功能模块。首先,利用集成学习中多个基学习器构建共识图;其次,利用图神经网络分析共识图,捕获节点优化表示和节点的聚类结构,并从聚类中挑选高置信度的节点子集及对应的类标签生成监督信息;再次,在此标签监督下,联合其他无标注样本更新集成成员基学习器。交替迭代上述功能块,最终提高无监督聚类的性能。结果为验证该框架的有效性,在标准数据集(包括图像和文本数据)上设计了一系列实验。实验结果表明,所提方法在性能上一致优于现有聚类方法。尤其是在MNIST-Test(modified national institute of standards and technology database)上,本文方法实现了97.78%的准确率,比已有最佳方法高出3.85%。结论该方法旨在利用图表示学习提升自监督学习中监督信息捕获的能力,监督信息的有效获取进一步强化了集成学习中成员构建的能力,最终提升了无监督海量数据本质结构的挖掘性能。展开更多
基金the Anhui Provincial Key Research and Development Plan(No.202104a06020012 and No.202204c06020022)the Major Natural Science Research Project of Universities in Anhui Province(No.2022AH040125).
文摘In crop image recognition,when faced with a large quantity of unlabeled data,the traditional manual labeling method requires a large amount of human and material resources.To solve this problem,this study proposes an image recognition method based on a pseudolabeling technique.First,the data are divided into labeled and unlabeled data.The initial network model is trained on labeled data.Then,pseudolabeling of the unlabeled data is predicted,and only the data that satisfy the confidence threshold are regarded as valid pseudolabeling.To convert the unlabeled data into supervised training data,the two types of data are mixed.The training is terminated when the number of remaining unlabeled data satisfies the end condition and when the fivefold crossvalidation method is used to evaluate model performance.Compared with the traditional semisupervised method,the experimental method is simpler and more applicable.Experiments were conducted on rice growth stage recognition and crop weed seedling recognition tasks.The results showed that the proposed method achieved 99.17%accuracy in rice growth stage recognition and a high AUC value of 99.93%in crop weed seedling recognition,which demonstrated excellent performance.Compared with the traditional model,this method not only improves in accuracy but also has better stability and wider applicability and is expected to provide an efficient,accurate and scalable solution for crop image recognition.
文摘目的随着实际应用场景中海量数据采集技术的发展和数据标注成本的不断增加,自监督学习成为海量数据分析的一个重要策略。然而,如何从海量数据中抽取有用的监督信息,并该监督信息下开展有效的学习仍然是制约该方向发展的研究难点。为此,提出了一个基于共识图学习的自监督集成聚类框架。方法框架主要包括3个功能模块。首先,利用集成学习中多个基学习器构建共识图;其次,利用图神经网络分析共识图,捕获节点优化表示和节点的聚类结构,并从聚类中挑选高置信度的节点子集及对应的类标签生成监督信息;再次,在此标签监督下,联合其他无标注样本更新集成成员基学习器。交替迭代上述功能块,最终提高无监督聚类的性能。结果为验证该框架的有效性,在标准数据集(包括图像和文本数据)上设计了一系列实验。实验结果表明,所提方法在性能上一致优于现有聚类方法。尤其是在MNIST-Test(modified national institute of standards and technology database)上,本文方法实现了97.78%的准确率,比已有最佳方法高出3.85%。结论该方法旨在利用图表示学习提升自监督学习中监督信息捕获的能力,监督信息的有效获取进一步强化了集成学习中成员构建的能力,最终提升了无监督海量数据本质结构的挖掘性能。
