类激活映射指导数据增强的细粒度图像分类  

Class Activation Mapping Guided Data Augmentation for Fine-Grained Visual Classification

作　　者：郭文明 王腾亿 Guo Wenming;Wang Tengyi(School of Computer Science(National Pilot Software Engineering School),Beijing University of Posts and Telecommunications,Beijing 100876;School of Information Engineering,Xinjiang Institute of Engineering,Urumqi 830023)

机构地区：[1]北京邮电大学计算机学院(国家示范性软件学院),北京100876 [2]新疆工程学院信息工程学院,乌鲁木齐830023

出　　处：《计算机辅助设计与图形学学报》2021年第11期1698-1704,共7页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(62162060)。

摘　　要：细粒度图像分类任务的关键在于获取精细的局部特征,为了充分利用数据价值,提出一种面向视觉注意力的数据增强方法,基于类激活映射图(class activation mapping,CAM)生成具有针对性的扩充图像,进而帮助细粒度分类.根据CAM对输入图像进行注意力区域裁剪和放大;构造一个流场网格对原图进行采样以夸张该区域,裁剪与夸张后的2种扩充数据能够引导模型学习更细微的特征差异;遮挡图像关键区域,从而促使模型学习其他有效特征.该方法只需要图像级标签,无需边界框和部位标注,可以在不引入其他辅助网络的情况下直接进行端到端训练.在3个公开数据集CUB-200-2011,FGVC-Aircraft和Stanford Cars上的实验结果表明,模型特征提取能力得到有效提升,且Top-1准确率指标优于部分现有先进算法.The key of fine-grained image classification is to extract discriminative partial features.In order to make full use of data,a visual attention guided data augmentation method is proposed to generate targeted im-ages based on Class Activation Mapping(CAM).Attention area found by CAM will be cropped and enlarged.A flow field grid will be generated to guide the sampling of original image so that the discriminative area can be exaggerated,therefore the network can learn more subtle features from two types of augmented images.An image with discriminative area dropped will be generated to encourage the network to learn other effective features.The algorithm only needs image-level labels without bounding boxes or parts labeling and can be trained end-to-end without other auxiliary networks.The experiments on CUB-200-2011,FGVC-Aircraft and Stanford Cars datasets demonstrate that capability of model is effectively improved,and the Top-1 accuracy metric is better than certain existing advanced algorithms.

关 键 词：细粒度图像分类 数据增强 类激活映射 视觉注意力 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]