北京航空航天大学学报 ›› 2019, Vol. 45 ›› Issue (9): 1864-1873.doi: 10.13700/j.bh.1001-5965.2018.0766

• 论文 • 上一篇    下一篇

基于深度卷积的残差三生网络研究与应用

厉铮泽, 杨小远, 朱日东, 王敬凯   

  1. 北京航空航天大学 数学与系统科学学院, 北京 100083
  • 收稿日期:2018-12-26 出版日期:2019-09-20 发布日期:2019-09-29
  • 通讯作者: 杨小远 E-mail:xiaoyuanyang@vip.163.com
  • 作者简介:厉铮泽,男,硕士研究生。主要研究方向:深度学习与应用;杨小远,女,博士,教授,博士生导师。主要研究方向:深度学习与图像处理;朱日东,男,博士研究生。主要研究方向:深度学习与目标跟踪;王敬凯,男,博士研究生。主要研究方向:深度学习与图像融合。
  • 基金资助:
    国家自然科学基金(61671002)

Research and application of residual triplet network based on deep convolution

LI Zhengze, YANG Xiaoyuan, ZHU Ridong, WANG Jingkai   

  1. School of Mathematics and Systems Science, Beihang University, Beijing 100083, China
  • Received:2018-12-26 Online:2019-09-20 Published:2019-09-29
  • Supported by:
    National Natural Science Foundation of China (61671002)

摘要: 针对图像多分类任务,提出基于深度卷积的残差三生网络,旨在通过残差学习和距离比较来训练神经网络得到有效的特征表示。首先,设计了一个21层的深度卷积神经网络作为三生网络的嵌入网络,其中该卷积网络共连接6个块(block)。利用残差学习的方式,每个block的输出层由卷积层的输出和该block的输入共同组成,降低网络学习难度,避免网络出现退化问题。然后,每个block中采用相同拓扑结构分路的卷积层,拓宽网络的宽度。最后,在全连接层拼接了来自前面卷积层和block的输出,加强特征信息的传递。训练前,针对正负样本采用交叉组合的采样方法来增加有效训练样本量;训练期间,用样本中心点更换原点样本作为输入,能平均降低0.5%错误率。在与其他三生网络的对比实验中,在MNIST、CIFAR10和SVHN数据库上达到最好的效果,在所有分类网络中,本文网络在MNIST上达到最好的效果,在CIFAR10和SVHN上表现优异。

关键词: 卷积神经网络, 三生损失, 残差学习, 挑战性样本采样, 样本中心点

Abstract: For multi-classification image tasks, a residual triplet network based on deep convolution is proposed, which aims to train neural networks to obtain useful feature representations through residual learning and distance comparison. Firstly, a 21-layer deep convolution neural network is designed as the embedded network of the triplet network, where the convolutional network is connected with 6 blocks. By using residual learning, the output of each block is combined with the input of this block and the output of the convolutional layer which focus on reducing the difficulty of network learning and avoiding degradation. Then, each block employed the convolution layers with the same topological branch to broaden the width of the network.Finally, to enhance the transfer of feature information, the fully-connected layer concatenated the output of the previous convolutional layers and blocks. Before training, the cross-combined sampling method is used to increase effective samples for hard samples. During training, using the sample center point to replace the anchor sample as an input can reduce the error rate by 0.5% on average. Among the triplet network series, we achieved the best results on the MNIST, CIFAR10, and SVHN. In all classification networks, we achieved the best results on the MNIST and performed well on CIFAR10 and SVHN.

Key words: convolution neural network, triplet loss, residual learning, hard sample mining, sample center point

中图分类号: 


版权所有 © 《北京航空航天大学学报》编辑部
通讯地址:北京市海淀区学院路37号 北京航空航天大学学报编辑部 邮编:100191 E-mail:jbuaa@buaa.edu.cn
本系统由北京玛格泰克科技发展有限公司设计开发