摘要
现有的基于注意力机制的细粒度图像识别方法大多都没有考虑目标局部的相关性,而且以往大多数方法都用多阶段或者多尺度机制,导致效率不高且难以端到端训练。本文提出的方法能调节不同输入图像的不同部位的关系。基于上述思路的注意力机制的方法去学习每幅图的每个关注区域特征,再用增强多重注意力机制强化这一效果,让同类别图像具有类似的注意力机制,而不同类别的图像具有不一样的注意力机制,同时也能够进行端到端训练。
Most of the existing fine-grained image recognition methods based on attention mechanism do not consider the local correlation of the target.In addition,most of the previous methods use multi-stage or multi-scale mechanism,which leads to low efficiency and difficulty in end-to-end training.This paper proposes that the relationship between different parts of different input images can be adjusted.The method based on the attention mechanism of the above ideas is to learn the characteristics of each focus area of each graph.Then the amplified multi-attention method is used to enhance the effect,so that the same category of images have similar attention mechanism,and different categories of images have different attention mechanism and can also be trained end-to-end.
作者
周晨轶
冯宇
徐亦白
卢杉
ZHOU Chen-yi;FENG Yu;XU Yi-bai;LU Shan(Information and Communication Branch,State Grid Zhejiang Electric Power Co.,Ltd.,Hangzhou 310012,China)
出处
《计算机与现代化》
2019年第9期83-89,共7页
Computer and Modernization
关键词
多注意力机制
端到端
细粒度图像识别
multiple attention mechanism
end-to-end
fine-grained image recognition
作者简介
周晨轶(1993-),男,浙江杭州人,助理工程师,硕士,研究方向:信息化项目管理,E-mail:macarzhou@163.com;冯宇(1982-),男,工程师,本科,研究方向:办公自动化业务及数据交换等平台类信息系统运维;徐亦白(1992-),男,助理工程师,硕士,研究方向:信息化项目管理;卢杉(1992-),男,助理工程师,硕士,研究方向:信息化项目管理。
