Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (2): 235-243.DOI: 10.3778/j.issn.1002-8331.2008-0402

• Graphics and Image Processing • Previous Articles     Next Articles

Weakly Supervised Fine-Grained Image Classification Based on Xception Network

DING Wenqian, YU Pengfei, LI Haiyan, LU Xinwei   

  1. School of Information Science and Engineering, Yunnan University, Kunming 650500, China
  • Online:2022-01-15 Published:2022-01-18

基于Xception网络的弱监督细粒度图像分类

丁文谦,余鹏飞,李海燕,陆鑫伟   

  1. 云南大学 信息学院,昆明 650500

Abstract: With the rapid development of deep learning, the classification of images research in the field of computer vision is not only limited to recognizing the categories of objects, but also needs more detailed classification based on the traditional image classification task. Based on the existing fine-grained image classification algorithm and model analysis, a model based on Xception model and WSDAN(weakly supervised data augmentation network) weak supervision data augment method of combination of deep learning network is applied to fine-grained image classification task. The method takes Xception network as the backbone network and feature extraction network, uses the improved WSDAN model for data augment, and feeds the augmented image back to the network as the input image to enhance the generalization ability of the network. Experiments on the commonly used fine-grained image data sets and NABirds data set show that the classification accuracy rate is 89.28%, 91.18%, 94.47%, 93.04% and 88.4%, respectively. The experimental results show that this method achieves better classification results compared with WSDAN(Pytorch) model and other mainstream fine-grained classification algorithms.

Key words: fine-grained image classification, data augment, deep learning, weak supervision, Xception network

摘要: 随着深度学习的快速发展,计算机视觉领域对图像的分类研究不仅仅局限于识别出物体的类别,更需要在传统图像分类任务的基础上进行更细致的类别划分。通过对现有细粒度图像分类算法和模型的分析研究,提出一种基于Xception模型与WSDAN(weakly supervised data augmentation network)弱监督数据增强的方法相结合的深度学习网络应用于细粒度图像分类任务。该方法以Xception网络作为骨干网络和特征提取网络、利用改进的WSDAN模型进行数据增强,并把增强后的图像反馈回网络作为输入图像来增强网络的泛化能力。在常用的细粒度图像数据集和NABirds数据集上进行实验验证,得到的分类正确率分别为89.28%、91.18%、94.47%、93.04%和88.4%。实验结果表明,与WSDAN(Pytorch)模型及其他多个主流细粒度分类算法相比,该方法取得了更好的分类结果。

关键词: 细粒度图像分类, 数据增强, 深度学习, 弱监督, Xception网络