2025年01期 v.46 59-65页
曹增辉1 陈浩1 曹雅慧2
（1.广东工业大学，广东 广州 510000
2.郑州工业安全职业学院，河南 郑州 450000）
摘要：小样本学习是图像分类任务中的一个重要挑战，能够有效解决因数据量较少而产生的模型准确率降低的问题。针对小样本学习难以准确获取类内共有特征的问题，提出一种基于类注意力的原型网络改进方法。利用掩膜图像进行数据预处理和图像增强，以提高原始数据质量；引入注意力机制，选择性地关注特征图中的重要信息，以增强特征提取能力；设计类注意力模块，提取具有注意力信息的类别原型。实验结果表明，在miniImageNet数据集上，该方法的分类准确率在基线基础上提高了2%，验证了其有效性。
关键词：原型网络；小样本学习；数据增强；类注意力；图像分类
中图分类号：TP183 文献标志码：A 文章编号：1674-2605(2025)01-0009-07
DOI：10.3969/j.issn.1674-2605.2025.01.009 开放获取
Improvement Method of Prototype Network Based on Class Attention
CAO Zenghui1 CHEN Hao1 CAO Yahui2
(1.Guangdong University of Technology, Guangzhou 510000, China
2.Zhengzhou Vocational College of Industrial Safety, Zhengzhou 450000, China)
Abstract: Small sample learning is an important challenge in image classification tasks, which can effectively solve the problem of reduced model accuracy due to limited data volume. A prototype network improvement method based on class attention is proposed to address the problem of difficulty in accurately obtaining common features within classes in small sample learning. Using mask images for data preprocessing and image enhancement to improve the quality of raw data; Introducing attention mechanism to selectively focus on important information in feature maps to enhance feature extraction capability; Design a class attention module to extract class prototypes with attention information. The experimental results show that on the miniImageNet dataset, the classification accuracy of this method has improved by 2% compared to the baseline, verifying its effectiveness.
Keywords: prototype network; small sample learning; data enhancement; class attention; image classification
