基于主动学习的图像分类技术: 现状与未来

Translated title of the contribution: Active Learning-Based Image Classification Technology: Status and Future

Ying Liu, Yu Liang Pang, Wei Dong Zhang, Da Xiang Li, Zhijie Xu

Research output: Contribution to journalArticlepeer-review

Abstract

As one of the important research directions in the field of computer vision, image classification has a wide range of applications. The success of deep learning-based image classification techniques depends on a large amount of annotated data. However, the cost of data annotation is often expensive. Active learning is a machine learning method that aims to achieve the expected model performance with as few high-quality annotated data as possible, and it can alleviate the problem of high annotation costs and difficulty in obtaining a large amount of annotation information in supervised learning tasks. Based on a sample selection strategy, active learning for image classification selects samples from the unlabelled dataset which are informative and thus contribute more to the training of the classification model, in order to update the annotated training data pool. This process is repeated until a given stopping condition is met or the model annotation budget is exhausted. This paper provides a comprehensive survey of the active learning image classification algorithms published in recent years. According the strategies applied in sample data processing and model structure optimization, existing algorithms are classified into three categories: algorithms based on data augmentation, including those using image augmentation to expand the scale of training data or using the differences in image feature interpolation to select high-quality training data; algorithms based on data distribution information, which optimize sample selection strategies based on the characteristics of data distribution; algorithms for optimizing model predictions, including methods for optimizing the acquisition and utilization of deep model prediction information, improving the predictive model structure through the use of generative adversarial networks and reinforcement learning, as well as enhancing model prediction performance based on the Transformer architecture to ensure the reliability of model predictions. In addition, this paper also conducts experimental comparisons on important academic work under various types of active learning image classification algorithms, and analyzes the performance and adaptability of each algorithm on datasets of different scales. Furthermore, this paper discusses the challenges faced by active learning image classification technology and points out future research directions.

图像分类作为计算机视觉领域中的重要研究方向之一,应用领域非常广泛. 基于深度学习的图像分类 技术取得的成功,依赖大量的已标注数据,然而数据的标注成本往往是昂贵的. 主动学习作为一种机器学习方法,旨 在以尽可能少的高质量标注数据达到期望的模型性能,缓解监督学习任务中存在的标注成本高、标注信息难以大量获 取的问题. 主动学习图像分类算法根据样本选择策略,从未标记样本数据集合中选择出信息量丰富,对分类模型训练 贡献更高的样本进行标注,以更新已标注训练数据池,如此循环直至满足给定的停止条件或模型标注预算耗尽. 本文 对近年来提出的主动学习图像分类算法进行了详细综述,并根据所用样本数据处理及模型优化方案,将现有算法分为 三类:基于数据增强的算法,包括利用图像增广来扩充训练数据,或者根据图像特征插值后的差异性来选择高质量的 训练数据;基于数据分布信息的算法,根据数据分布的特点来优化样本选择策略;优化模型预测的算法,包括优化获取 和利用深度模型预测信息的方法、基于生成对抗网络和强化学习来优化预测模型的结构,以及基于Transformer结构提 升模型预测性能,以确保模型预测结果的可靠性. 此外,本文还对各类主动学习图像分类算法下的重要学术工作进行 了实验对比,并对各算法在不同规模数据集上的性能和适应性进行了分析. 另外,本文探讨了主动学习图像分类技术 所面临的挑战,并指出了未来研究的方向.

Translated title of the contributionActive Learning-Based Image Classification Technology: Status and Future
Original languageChinese (Traditional)
Pages (from-to)2960-2984
Number of pages25
JournalTien Tzu Hsueh Pao/Acta Electronica Sinica
Volume51
Issue number10
Early online date25 Oct 2023
Publication statusPublished - 7 Dec 2023

Cite this