小样本目标检测综述

A Survey of Few-shot Object Detection

小样本学习是指在样本数据不足或质量较低的情况下进行的深度学习训练和预测的方法。针对深度学习目标检测应用中可能会面对的样本数据不足的问题,分析了小样本目标检测的数学模型和误差来源,将适用于小样本目标检测的方法分成数据、模型和算法三个类别进行了归纳总结,简述了各个方案的缺点与不足,并枚举了近年来在小样本目标检测上的可行方法实践探索,简要介绍了其实现的效果。在此基础上,简单介绍了与小样本学习相类似的深度学习应用,并在分析了目前小样本检测中存在的问题后,对未来小样本目标检测的发展方向和研究趋势进行了讨论。

Few-shot learning refers to the method of deep learning training and prediction under the condition of insufficient or low-quality sample data.Aiming at the problem of insufficient sample data that may be faced in the application of deep learning object detection,the mathematical model and error source of few-shot object detection are analyzed firstly.The methods applicable to few-shot object detection are given in three categories:data,model and algorithm,and the shortcomings of each scheme are attached.Based on the practical exploration of few-shot object detection,the recent attempts along with their results are enumerated.The other applications of deep learning similar to few-shot learning are also briefly introduced.Then,after analyzing the existing problems in few-shot detection,the development direction and research trend of few-shot object detection in the future are discussed.