计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (24): 223-232.DOI: 10.3778/j.issn.1002-8331.2105-0290

• 图形图像处理 • 上一篇    下一篇

基于自适应特征融合与转换的小样本图像分类

许栋,杨关刘小明,刘阳,刘济宗,陈静,郭清宇   

  1. 1.中原工学院 计算机学院,郑州 450007
    2.河南省网络舆情监测与智能分析重点实验室,郑州 450007
    3.中原工学院 前沿信息技术研究院 网络舆情研究中心,郑州 450007
    4.西安电子科技大学 通信工程学院,西安 710071
  • 出版日期:2022-12-15 发布日期:2022-12-15

Few-Shot Learning Image Classification Based on Adaptive Feature Fusion and Transformation

XU Dong, YANG Guan, LIU Xiaoming, LIU Yang, LIU Jizong, CHEN Jing, GUO Qingyu   

  1. 1.School of Computer Science, Zhongyuan University of Technology, Zhengzhou 450007, China
    2.Henan Key Laboratory on Public Opinion Intelligent Analysis, Zhengzhou 450007, China
    3.Internet Public Opinion Research Center, the Frontier Information Technology Research Institute, Zhongyuan University of Technology, Zhengzhou 450007, China
    4.School of Telecommunications Engineering, Xidian University, Xi’an 710071, China
  • Online:2022-12-15 Published:2022-12-15

摘要: 小样本学习中数据采样不断变化的特点使得模型特征提取不充分,同时,模型对提取的特征也难以进行相应操作;数据分布的变化也影响着小样本模型的性能。针对这些问题,提出一种基于自适应加权多路分支小样本图像分类模型。多路特征处理模块对输入数据进行特征提取和融合,以便充分利用少量数据;自适应的支路权重使得特征信号随特征进行相应的放缩;特征转换模块对多变的数据分布进行适应性变化,以便更好地聚合同类,提高分类效果。通过使用Caltech-UCSD Birds-200-2011数据集和mini-ImageNet数据集,对所提模型在不同场景下进行分类效果测试。实验结果表明,所提模型在5-Way 1-Shot和5-Way 5-Shot任务中的准确率分别比baseline相比分别提升9.81、8.16个百分点和9.16、9.21个百分点,验证了模型的有效性。

关键词: 图像分类, 小样本学习, 特征融合, 特征转换, 自适应加权多路分支

Abstract: The constant variation of data sampling in few shot learning makes the feature extraction of the model inadequate, and it is difficult for the model to manipulate the extracted features. Furthermore, the variation in the data distribution also affects the performance of the few shot model. To address these problems, a few shot image classification model based on adaptive weighted multi-branch is proposed. The multi-channel feature processing module performs the feature extraction and fusion on the input data so that a small data can be fully utilized, and the adaptive branch weights make the feature signals scale with the features. In addition, the feature conversion module adaptively changes to variable data distributions to better aggregate the same classes and improve the classification performance. The proposed model is tested on the Caltech-UCSD Birds-200-2011 dataset and mini-ImageNet dataset for classification performance on different scenes. The classification performance of the proposed model is tested on the Caltech-UCSD Birds-200-2011 dataset and the mini-ImageNet dataset under different scenarios. The experimental results demonstrate that the accuracy of the proposed model is improved by 9.81, 8.16 percentage points and 9.16, 9.21 percentage points in the 5-Way 1-Shot and 5-Way 5-Shot tasks, respectively, compared with baseline. This also shows the effectiveness of the model.

Key words: image classification, few shot learning, feature fusion, feature conversion, adaptive weighted multi-branch