跨域图像空间数据少样本学习研究

Research on Cross-domain Image Spatial Data Few-shot Learning

跨域图像空间数据少样本学习(Few-Shot Learning, FSL)是近年来机器学习研究领域的热点,旨在利用少量的有标签图像空间源域数据训练一个可靠的模型对分布差异大的图像目标域数据进行分类。概述了近年来主要的跨域图像空间数据FSL模型,根据模型解决问题的主要思想,将其分类为数据引入法、特征增强法、参数控制法以及混合法。将数据引入法细分为基于单源域数据、基于多源域数据和基于目标域数据;将特征增强法细分为特征转换和特征融合;将混合法细分为不同方法的结合使用和不同类型损失函数的结合使用,并总结了不同方法的原理、优点与不足。对当前跨域图像空间FSL常用的数据集、基准进行了详细介绍,在主流基准上对经典模型的实验结果进行对比与分析。对当前跨域图像空间数据FSL面临的挑战进行总结,指出未来可能的发展方向。

Cross-domain image spatial data Few-Shot Learning(FSL)aims to train a reliable model using a small amount of labeled image spatial source domain data to classify image target domain data with large distribution differences.It is a hot topic in the field of machine learning research in recent years.An overview of the main current cross domain image spatial data FSL models is provided.Based on the main ideas of model problem-solving,the models are classified into data introduction method,feature enhancement method,parameter control method,and hybrid method.Among them,the data introduction is subdivided into single source domain data,multi-source domain data,and target domain data.The feature enhancement method is subdivided into feature transformation and feature fusion.The hybrid method is subdivided into the combination of different methods and the combination of different types of loss functions,the principles,advantages,and disadvantages of different methods are summarized.At the same time,a detailed introduction to the commonly used datasets and benchmarks for cross domain image space FSL is also provided,and the experimental results of classical models on mainstream benchmarks are compared and analyzed.Finally,the challenges of the crossdomain image spatial data learning with FSL is summarized,and the future research directions is pointed out.