特征数据增广探究

Expanded Exploration of Feature Spatial Data

深度学习模型的进步很大程度上归功于收集到的数据集,在数量和数据多样性的增加。数据增广是研究人员不用实际收集数据的情况下,能够显著提升训练模型性能的有效方法之一。诸如裁剪、填充和水平翻转等数据增广技术通常被用于训练大型神经网络。然而,训练神经网络的大多数方法仅仅使用了基本类型的数据增广技术。一种增广技术在不同数据集上并不能通用,并且耗时。目前研究人员,在特征空间寻找简单的数据增广方法,解决训练耗时,力求方法通用性,减少人工成本而努力。本文将在数据增广的框架下总结近年来各种数据增广方法,以便相关领域的研究人员更全面了解目前研究进展。最后对于在特征空间数据增广领域存在的问题和发展趋势进行总结和展望。

The advances in deep learning models are largely due to the increase in the number and diversity of data sets collected.Data augmentation is one of the effective methods for researchers to significantly improve the performance of training models without actually collecting data.Data augmentation techniques such as cropping,filling,and horizontal flipping are commonly used to train large neural networks.However,most methods for training neural networks use only basic types of data augmentation techniques.An augmentation technique is not universal across different data sets and is time consuming.At present,researchers are trying to find a simple method of data augmentation in the feature space,so as to solve the time-consuming training,strive for the universality of the method and reduce the labor cost.In this paper,various data augmentation methods in recent years will be summarized under the framework of data augmentation,so that researchers in related fields can have a more comprehensive understanding of the current research progress.Finally,the existing problems and development trends in the field of feature spatial data augmentation are summarized and prospected.