基于深度学习的无监督样本标注方法研究

Research on unsupervised sample labeling method based on deep learning

在早期的样本标记阶段,手工为样本做标注需要大量人力,既耗时又耗力。为实现更准确、高效的自动标注,文章提出了利用多模型进行协同标注的图像自动化标注(AutomaticImage Annotation,AIA)方法。为了验证AIA方法的普适性,文章选择了3个数据集(MNIST数据集、猫狗数据集、Cifar10数据集),并在每个数据集上训练了3个模型。使用Keras框架构建并训练这些模型,通过投票策略和加权投票策略进行样本标注。通过对比分析发现,加权投票策略效果最优。

In the early stage of sample labeling,manually labeling samples requires a lot of manpower,which is both time-consuming and labor-intensive.To achieve more accurate and efficient automatic annotation,the article proposes an Automatic Image Annotation(AIA)method that utilizes multiple models for collaborative annotation.To verify the universality of the AIA method,the article selected three datasets(MNIST dataset,cat and dog dataset,Cifar10 dataset)and trained three models on each dataset.Build and train these models using the Keras framework,and annotate samples through voting strategies and weighted voting strategies.Through comparative analysis,it was found that the weighted voting strategy has the best effect.