计算机工程 ›› 2019, Vol. 45 ›› Issue (5): 205-209.doi: 10.19678/j.issn.1000-3428.0051191

• 图形图像处理 • 上一篇    下一篇

递归深度混合关注网络的细粒度图像分类方法

桂江生1,麻陈飞1,包晓安1,钱俊彦2   

  1. 1.浙江理工大学 信息学院,杭州 310018; 2.桂林电子科技大学 广西可信软件重点实验室,广西 桂林 541004
  • 收稿日期:2018-04-12 出版日期:2019-05-15 发布日期:2019-05-15
  • 作者简介:桂江生(1978—),男,副教授,主研方向为计算机视觉、模式识别;麻陈飞,硕士研究生;包晓安、钱俊彦,教授。
  • 基金项目:

    国家自然科学基金(61502430,61562015);浙江省重大科技专项重点工业项目(2014C01047);广西自然科学重点基金(2015GXNSFDA139038)。

Fine-grained Image Classification Method for Recurrent Deep Hybrid Attention Network

GUI Jiangsheng1,MA Chenfei1,BAO Xiaoan1,QIAN Junyan2   

  1. 1.School of Informatics Science and Technology,Zhejiang Sci-tech University,Hangzhou 310018,China; 2.Guangxi Key Laboratory of Trusted Software,Guilin University of Electronic Technology,Guilin,Guangxi 541004,China
  • Received:2018-04-12 Online:2019-05-15 Published:2019-05-15

摘要:

在细粒度图像的大量局部特征中,只有少量特征具有判别性,其提取较为困难。为此,提出递归深度混合关注网络方法。通过在卷积结构单元中添加通道关注模块和空间关注模块,实现网络的混合关注。以第1路网络输出特征的空间响应值为依据切割原图,并将切割后的图像放大输入第2路网络,进行由粗到细的网络递归。将2路网络提取的特征进行级联融合。在公开数据集Stanford Dogs、Stanford Cars中进行对比实验,结果表明,该方法的分类精度分别为87.1%、92.4%,优于FCAN、HIHCA等方法。

关键词: 细粒度图像分类, 卷积神经网络, 通道关注, 空间关注, 递归网络

Abstract:

Among the large number of local features of fine-grained images,only a few of them are discriminative and are very difficult to extract.To solve this problem,a recurrent deep hybrid attention network is proposed.By adding channel attention module and spatial attention module into convolution units,the hybrid attention network is realized.Based on the spatial response value of the output feature of the first network,the original image is segmented,and the segmented image is enlarged and input into the second network to realize the coarse-to-fine network recursion.The features extracted from two networks are cascaded to achieve the purpose of feature fusion.Experiments are conducted in the open data sets.Stanford Dogs and Stanford Cars.Results show that the classification accuracies of the proposed algorithm are 87.1% and 92.4%,respectively,which is better than FCAN,HIHCA and other methods.

Key words: fine-grained image classification, convolutional neural network, channel attention, spatial attention, recurrent network

中图分类号: