基于改进深度子域适应网络的图像分类方法
Image Classification Method Based on Improved Deep Subdomain Adaptation Network
郝海燕 1李芳1
作者信息
- 1. 沈阳理工大学 信息科学与工程学院,沈阳 110159
- 折叠
摘要
深度子域适应网络在进行特征提取时易导致部分重要信息丢失,且在对齐局部特征的同时会忽略源域和目标域的整体对齐,从而影响其分类准确率,为此提出基于改进深度子域适应网络的图像分类方法.引入卷积神经网络正则化方法提升模型在不同跨域任务中的泛化能力;在特征提取网络中加入高效通道注意力机制,对局部跨信道交互信息进行捕捉,提取输入图像中的关键信息;改进损失函数,增加全域适应损失约束,提升模型的全局特征对齐效果.在域适应基准数据集Office-31 上的实验表明,相较于原算法,本文改进算法在一定程度上提升了分类准确率,在跨域图像分类任务中表现更好.
Abstract
The deep subdomain adaptation network easily leads to the loss of some important infor-mation during feature extraction,and ignores the overall alignment of the source domain and the tar-get domain while aligning local features,thus affecting the classification accuracy.To address these issues,an improved method based on the deep subdomain adaptation network is proposed.The con-volutional neural network regularization method is introduced to improve the model's generaliza-tion ability in different cross-domain tasks.Incorporating an efficient channel attention mechanism into the feature extraction network can extract key information from the input image by capturing local cross-channel interaction information.The loss function is modified by adding a global domain adaptation loss constraint which enhances the overall feature alignment effect of the model.Experi-ments on the domain-adapted benchmark dataset Office-31 show that the improved algorithm im-proves the classification accuracy to a certain extent compared with the original algorithm,and per-forms better in cross-domain image classification tasks.
关键词
图像分类/子域适应/高效通道注意力/全局特征对齐Key words
image classification/subdomain adaptation/efficient channel attention/global feature a-lignment引用本文复制引用
出版年
2024
NETL
NSTL
万方数据