融合多粒度注意力特征的小样本分类模型

Few-shot classification model incorporating multi-granular attention features

在小样本分类任务中,现有的CNN模型存在特征提取不足、特征单一和小样本数据集类间差异化较弱的问题,导致分类精度较低。针对以上问题,提出一种融合多粒度注意力特征(fusion multi-granular attention feature,FMAF)的小样本分类模型。首先,该方法借鉴多粒度思想,重新设计CNN特征提取网络的架构来增强特征多样性;其次,在多粒度特征提取网络后添加自注意力层,提取多粒度图像特征中的关键特征,在多粒度注意力特征的基础上,借助特征融合方法融合多粒度注意力特征信息,突出关键特征,提高特征的表征力;最后,在两个经典的小样本数据集miniImageNet和tieredImageNet上进行了评估。实验结果表明,FMAF方法能有效提升分类的准确度和效率。

In the few-shot classification tasks,existing CNN models suffer from insufficient feature extraction,limited feature diversity and weak differentiation between classes in few-shot datasets,leading to low classification accuracy.To address these issues,this paper proposed a few-shot classification model called FMAF.Firstly,this method incorporated multi-granularity thought into the architecture of CNN feature extraction network to enhance feature diversity.Secondly,after the multi-granular feature extraction network,FMAF added a self-attention layer to extract key features from the multi-granular image features,based on the multi-granular attention features,FMAF employed a feature fusion method to combine the information from multiple-granularity attention features,highlighted the crucial features and improved feature representativeness.Finally,this paper utilized two classical few-shot datasets for experimental verification on miniImageNet and tieredImageNet.Experimental results show that FMAF method can effectively improve the accuracy and efficiency of classification.