小样本困境下的深度学习图像识别综述

Survey on Deep Learning Image Recognition in Dilemma of Small Samples

图像识别是图像研究领域的核心问题,解决图像识别问题对人脸识别、自动驾驶、机器人等各领域研究都有重要意义.目前广泛使用的基于深度神经网络的机器学习方法,已经在鸟类分类、人脸识别、日常物品分类等图像识别数据集上达到了超过人类的水平,同时越来越多的工业界应用开始考虑基于深度神经网络的方法,以完成一系列图像识别业务.但是深度学习方法极度依赖大规模标注数据,这一缺陷极大地限制了深度学习方法在实际图像识别任务中的应用.针对这一问题,越来越多的研究者开始研究如何基于少量的图像识别标注样本来训练识别模型.为了更好地理解基于少量标注样本的图像识别问题,广泛地讨论了几种图像识别领域主流的少量标注学习方法,包括基于数据增强的方法、基于迁移学习的方法以及基于元学习的方法,通过讨论不同算法的流程以及核心思想,可以清晰地看到现有方法在解决少量标注的图像识别问题上的优点和不足.最后针对现有方法的局限性,指出了小样本图像识别未来的研究方向.

Present machine learning methods have reached a higher level than human intelligence in image recognition and other tasks.However,recent machine learning methods,especially deep learning methods,rely heavily on a large number of annotation data that human cognition often does not need.This weakness greatly limits the application of deep learning method in practical problem.To solve this problem,learning from a few shot examples attracts more and more community’s research interest.In order to better understand the few shot learning problem,this study extensively discusses several popular few shot learning methods,including data augmentation methods,transfer learning methods,and meta learning methods.After the processes and core ingredients of different algorithms are discussed,the advantages and disadvantages of existing methods could be clearly seen in solving few shot learning problems.At the end of this paper,the points to future research directions are highlighted in the field of few shot learning problem.