面向图像分类的小样本学习算法综述

Survey on Few-shot Learning Algorithms for Image Classification

目前,以深度学习为代表的人工智能算法凭借超大规模数据集以及强大的计算资源,在图像分类、生物特征识别、医疗辅助诊断等领域取得了优秀的成果并成功落地.然而,在许多实际的应用场景中,因诸多限制,研究人员无法获取到大量样本或者获取样本的代价过高,因此研究图像分类任务在小样本情形下的学习算法成为了推动智能化进程的核心动力,同时也成为了当下的研究热点.小样本学习指在监督信息数量有限的情况下进行学习并解决问题的算法.首先,从机器学习理论的角度描述了小样本学习困难的原因;其次,根据小样本学习算法的设计动机将现有算法归为表征学习、数据扩充、学习策略三大类,并分析其优缺点;然后,总结了常用的小样本学习评价方法以及现有模型在公用数据集上的表现;最后,讨论了小样本图像分类技术的难点及未来的研究趋势,为今后的研究提供参考.

Presently,artificial intelligence algorithms represented by deep learning have achieved advanced results and been successfully used in fields such as image classification,biometric recognition and medical assisted diagnosis by virtue of ultra-large-scale data sets and powerful computing resources.However,due to many restrictions in the actual environment,it is impossible to obtain a large number of samples or the cost of obtaining samples is too high.Therefore,studying the learning algorithm in the case of small samples is the core driving force to promote the intelligent process,and it has also become a current research hot-spot.Few-shot learning is the algorithm to learn and solve the problem under the condition of limited supervision information.Firstly,it describes the reasons why few-shot learning is difficult to generalize from the perspective of machine learning theory.Secondly,according to the design motivation of the few-shot learning algorithm,existing algorithms are classified into three categories:representation learning,data expansion and learning strategy,and their advantages and disadvantages are analyzed.Thirdly,we summarize the commonly used few-shot learning evaluation methods and the performance of existing models in public data sets.Finally,we discuss the difficulties and future research trends of small sample image classification technology to provide re-ferences for future research.