首页|基于注意力机制和迁移学习的服装分类方法

基于注意力机制和迁移学习的服装分类方法

扫码查看
针对服装图像分类效率低、准确率不高等问题,提出了一种基于注意力机制和迁移学习的服装图像分类方法.主要采用预训练的ResNet50网络模型在服装数据集上进行迁移学习,以降低对数据集的依赖,并减少网络训练时间;通过几何变换和颜色抖动2种数据增强手段处理数据集,提高模型的泛化能力;以ResNet50为基础网络,加入卷积注意力机制模块(convolutional block at-tention module,CBAM),依次从通道和空间2个维度提高对服装不同区域的关注度,增强了特征表达能力.在CD和IDFashion两类背景干扰信息不同的数据集上进行验证,实验结果表明:所提出的模型能够提取更多服装特征信息,在IDFashion数据集的平均分类准确率为95.60%,分别高于ResNet50、ResNet50+STN和ResNet50+ECA模型6.65%、6.69%、6.62%,一定程度上提高了服装图像分类的准确率和效率.
Clothing classification method based on attention mechanism and transfer learning
Aimed the low efficiency and low accuracy of clothing image classification,a clothing image classification method based on attention mechanism and transfer learning was proposed.The pre-trained ResNet50 network model was used for transfer learning on the clothing dataset to reduce the dependence on the dataset and the network training time.Image dataset was processed by data augmentation of geometric transform and color jitter to improve the generalization ability of the model.Convolutional block attention module(CBAM)was added to the ResNet50-based network,and attention of different region of clothing was improved from both channel and spatial dimensions in turn.Then the feature expression capability was enhanced.The validation was per-formed on two datasets of CD and IDFashion with different background interference.Experimen-tal results show that the proposed model can extract more clothing feature information,and the average classification accuracy in the IDFashion dataset is 95.60%,which is higher than that of ResNet50,ResNet50+STN and ResNet50+ECA models by 6.65%,6.69%,6.62%,which improves the accuracy and efficiency of clothing image classification to some extent.

clothing classificationResNet50convolutional block attention moduleattention mechanismtransfer learning

陈金广、黄晓菊、马丽丽

展开 >

西安工程大学 计算机科学学院,陕西 西安 710048

服装图像分类 ResNet50 卷积注意力机制模块(CBAM) 注意力机制 迁移学习

陕西省自然科学基础研究计划陕西省教育厅科研项目

2023-JC-YB-56822JP028

2024

西安工程大学学报
西安工程大学

西安工程大学学报

CSTPCD
影响因子:0.473
ISSN:1674-649X
年,卷(期):2024.38(3)
  • 10