摘要
针对在基于卷积神经网络的图像处理领域内,大部分特征融合只是通过Add或者Concat操作进行特征叠加或特征拼接而不能很好地将有效特征进行融合的问题,对Add和Concat特征融合引入通道域的注意力机制,设计了4种可学习的特征融合方式:A-Cat、B-Cat、A-Add和B-Add.为了验证方法的有效性,选择YOLOv3-Tiny作为baseline,在Pascal VOC2007数据集上进行测试.结果表明:A-Cat比原Concat的mAP提高了0.76%,比B-Cat提高了1.49%;A-Add比原Add的mAP提高了0.34%,比B-Add提高了1.41%.基于注意力机制的特征融合方式可以通过学习不同特征的重要程度并据此进行特征融合,有效地提升网络的性能.
Feature fusion is widely used during image processing based on convolution neural network.Most methods only use add or Concat to fuse features,and they can not be fused effectively.Therefore,the channel domain attention mechanism is introduced into the feature fusion of Add and Concat,and four learning feature fusion methods are designed:A-Cat,B-Cat,A-Add and B-Add.In order to verify the validity of the method,YOLOv3-Tiny was selected to be baseline,and it was tested on Pascal VOC2007 data set.The test results show that the mAP of A-Cat is improved 0.76%,B-Cat is improved 1.49%,A-Add is improved 0.34%and B-Add is improved 1.41%.Experiments show that feature fusion based on attention mechanism can effectively improve network performance by learning the importance of different features and fusing them accordingly.
作者
罗大为
方建军
刘艳霞
LUO Da-wei;FANG Jian-jun;LIU Yan-xia(Beijing Key Laboratory of Information Service Engineering,Beijing Union University,Beijing 100101,China;College of Urban Rail Transit and Logistics,Beijing Union University,Beijing 100101,China)
出处
《东北师大学报(自然科学版)》
北大核心
2021年第3期44-48,共5页
Journal of Northeast Normal University(Natural Science Edition)
基金
国家自然科学基金资助项目(61602041)
北京联合大学人才强校优选计划项目(BPHR2017CZ07).