模式识别与人工智能
2025年1月12日 星期日   首 页     期刊简介     编委会     投稿指南     伦理声明     联系我们                                                                English
模式识别与人工智能  2019, Vol. 32 Issue (4): 336-344    DOI: 10.16451/j.cnki.issn1003-6059.201904006
论文与报告 最新目录| 下期目录| 过刊浏览| 高级检索 |
基于稀疏化双线性卷积神经网络的细粒度图像分类
马力1, 王永雄1,2
1.上海理工大学 光电信息与计算机工程学院 上海 200093
2.上海康复器械工程技术研究中心 上海 200093
Fine-Grained Visual Classification Based on Sparse Bilinear Convolutional Neural Network
MA Li1, WANG Yongxiong1,2
1.School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093
2.Shanghai Engineering Research Center of Assistive Devices, Shanghai 200093

全文: PDF (766 KB)   HTML (1 KB) 
输出: BibTeX | EndNote (RIS)      
摘要 针对双线性卷积神经网络(B-CNN)在细粒度图像分类中因参数过多、复杂度过高而导致的过拟合问题,提出稀疏化B-CNN.首先对B-CNN的每个特征通道引入比例因子,在训练中采用正则化方法对其稀疏.然后利用比例因子的大小判别特征通道的重要性.最后将不重要特征通道按一定比例裁剪,消除网络过拟合,提高关键特征的显著性.稀疏化B-CNN属于弱监督学习,可实现端到端训练.在FGVC-aircraft、Stanford dogs、Stanford cars这3个细粒度图像数据集上的实验表明,稀疏化B-CNN的准确率高于B-CNN,也优于或基本接近其它通用的细粒度图像分类算法.
服务
把本文推荐给朋友
加入我的书架
加入引用管理器
E-mail Alert
RSS
作者相关文章
马力
王永雄
关键词 细粒度图像分类双线性卷积神经网络(B-CNN)过拟合网络稀疏网络剪枝    
Abstract:The overfitting problem of bilinear convolutional neural network(B-CNN) for fine-grained visual recognition is caused by the large number of parameters and its complex structure. In this paper, a sparse B-CNN is proposed to handle the problem. Firstly, a scaling factor is introduced into each feature channel of B-CNN, and regularization of sparsity is applied to the scaling factors during the training. Then, the feature channels in B-CNN with low contribution to the final classification are identified by small scaling factors. Finally, these channels are pruned in a certain proportion to prevent overfitting and increase the significance of key features. The learning of sparse B-CNN is weakly supervised and end-to-end. The verification experiments on FGVC-aircraft, Stanford dogs and Stanford cars fine-grained image datasets show that the accuracy of sparse B-CNN is higher than that of the original B-CNN. Moreover, compared with other advanced algorithms for fine-grained visual recognition, the performance of sparse B-CNN is same or even better.


Key wordsFine-Grained Visual Recognition    Bilinear Convolutional Neural Network(B-CNN)   
Overfitting
   Network Sparsity    Network Pruning   
收稿日期: 2019-01-03     
ZTFLH: TP 391  
基金资助:国家自然科学基金项目(No.61673276,61703277)资助
作者简介: 马 力,硕士研究生,主要研究方向为计算机视觉、图像处理.E-mail:mali1906@163.com.王永雄(通讯作者),博士,教授,主要研究方向为智能机器人及视觉.E-mail:wyxiong@usst.edu.cn.
引用本文:   
马力, 王永雄. 基于稀疏化双线性卷积神经网络的细粒度图像分类[J]. 模式识别与人工智能, 2019, 32(4): 336-344. MA Li, WANG Yongxiong. Fine-Grained Visual Classification Based on Sparse Bilinear Convolutional Neural Network. , 2019, 32(4): 336-344.
链接本文:  
http://manu46.magtech.com.cn/Jweb_prai/CN/10.16451/j.cnki.issn1003-6059.201904006      或     http://manu46.magtech.com.cn/Jweb_prai/CN/Y2019/V32/I4/336
版权所有 © 《模式识别与人工智能》编辑部
地址:安微省合肥市蜀山湖路350号 电话:0551-65591176 传真:0551-65591176 Email:bjb@iim.ac.cn
本系统由北京玛格泰克科技发展有限公司设计开发 技术支持:support@magtech.com.cn