可微风格搜索:一种在线自动数据增强方法  

Differentiable Style Searching:An Online Automatic Data Augmentation Method

作　　者：罗亚威 于俊清[2] Luo Yawei;Yu Junqing(School of Software Technology,Zhejiang University,Ningbo 315048;School of Computer Science and Technology,Huazhong University of Science and Technology,Wuhan 430074)

机构地区：[1]浙江大学软件学院,宁波315048 [2]华中科技大学计算机科学与技术学院,武汉430074

出　　处：《计算机辅助设计与图形学学报》2023年第4期553-561,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(62206249);博士后中期特别资助基金(2022T150567)。

摘　　要：针对现有数据增强方法中所存在的离线图像变换和搜索空间受限等问题,提出一种在线自动数据增强(ODA)方法.ODA的核心是一个可微风格搜索模块,该模块可以通过直接回传训练损失的方式感知任务模型当前所需的数据增强,以对抗搜索的方式在线生成更难的风格化图片以扩展训练集,从而高效地帮助模型完成在多种未知风格上的泛化.在跨领域图像分类任务的MNIST,MNIST-M,SVHN和USPS数据集以及跨领域场景语义分割任务的Cityscapes和GTA5数据集上,与其他5种典型的数据增强方法进行对比实验表明,在Acc指标上ODA方法能带来至少2%的分类精度提升,而在语义分割任务的mIoU指标上ODA方法能带来3%到7%的提升,证明了ODA扩充了传统自动数据增强方法在图像风格方向上的搜索空间,增强了网络的泛化能力.Aiming at the problems of offline image transformation and limited search space in existing data enhancement methods,an online automatic data enhancement(ODA)method is proposed.The core of ODA is a micro-style search module,which can perceive the current data enhancement required by the task model by directly transmitting training losses,and generate more difficult stylized pictures online to expand the training set by means of anti-search,so as to effectively help the model complete the generalization on a variety of unknown styles.On the cross-domain image classification task of MNIST,MNIST-M,SVHN and USPS data sets,as well as the cross-domain scene semantic segmentation task of Cityscapes and GTA5 data sets,compared with other five typical data enhancement methods,ODA method can improve classification accuracy by at least 2%under Acc metric and 3%to 7%under mIoU metric of semantic segmentation task,which proves that ODA extends the search space of traditional automatic data enhancement methods in the direction of image style and enhances the generalization ability of network.

关 键 词：自动数据增强 风格迁移 模型泛化 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]