人工智能培训

搜索

人工智能论文:寻找细节中的魔鬼:学习三线注意采样网络进行细粒度图像识别(Looking for the Devil in the Details: Learni

[复制链接]
xbgzs2010 发表于 2019-3-15 12:47:36 | 显示全部楼层 |阅读模式
xbgzs2010 2019-3-15 12:47:36 510 0 显示全部楼层
人工智能论文:寻找细节中的魔鬼:学习三线注意采样网络进行细粒度图像识别(Looking for the Devil in the Details: Learning Trilinear Attention  Sampling Network for Fine-grained Image Recognition)学习细微但有辨别力的特征(例如,鸟的喙和眼睛)在细粒度图像识别中起着重要作用。基于现有注意力的方法定位和放大重要部分以学习细粒度细节,这些细节通常受到有限数量的部件和大量计算成本的影响。在本文中,我们建议通过Trilinear Attention SamplingNetwork(TASN)以高效的师生方式从数百个部分提案中学习这些细粒度特征。具体来说,TASN包括1)三线性注意模块,它通过模拟信道间关系产生注意力图,2)基于注意力的采样器,其突出显示高分辨率的零件; 3)特征蒸馏器,将部分特征提炼为全局一个按重量分享和特征保留策略。大量实验证明,TASN在最具竞争力的方法,iniNaturalist-2017,CUB-Bird和Stanford-Cars数据集的相同设置下产生最佳性能。
Learning subtle yet discriminative features (e.g., beak and eyes for a bird)plays a significant role in fine-grained image recognition.Existingattention-based approaches localize and amplify significant parts to learnfine-grained details, which often suffer from a limited number of parts andheavy computational cost.In this paper, we propose to learn such fine-grainedfeatures from hundreds of part proposals by Trilinear Attention SamplingNetwork (TASN) in an efficient teacher-student manner.Specifically, TASNconsists of 1) a trilinear attention module, which generates attention maps bymodeling the inter-channel relationships, 2) an attention-based sampler whichhighlights attended parts with high resolution, and 3) a feature distiller,which distills part features into a globalone by weight sharing and featurepreserving strategies.Extensive experiments verify that TASN yields the bestperformance under the same settings with the most competitive approaches, iniNaturalist-2017, CUB-Bird, and Stanford-Cars datasets.人工智能论文:寻找细节中的魔鬼:学习三线注意采样网络进行细粒度图像识别(Looking for the Devil in the Details: Learning Trilinear Attention  Sampling Network for Fine-grained Image Recognition) mqXX3lvTqd99d9lX.jpg
URL地址:https://arxiv.org/abs/1903.06150     ----pdf下载地址:https://arxiv.org/pdf/1903.06150    ----人工智能论文:寻找细节中的魔鬼:学习三线注意采样网络进行细粒度图像识别(Looking for the Devil in the Details: Learning Trilinear Attention  Sampling Network for Fine-grained Image Recognition)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

xbgzs2010当前离线
新手上路

查看:510 | 回复:0

快速回复 返回顶部 返回列表