Abstract: How to extract important features by existing metric-based few-shot image classification models is a difficulty. A few-shot image classification method based on local orthogonal feature fusion is proposed. First, the feature extraction network is used to simultaneously extract shallow features with rich local details and deep features with strong semantics. Then, a channel attention module and a multi-scale feature adaptive fusion module are used to perform feature enhancement on the channel and scale dimensions of the shallow features, respectively, in order to generate the feature with more salient local features and more scale information. Finally, according to local orthogonal feature extraction and attention fusion, the obtained multi-scale local features and initial deep semantic features are extracted and fused by a local orthogonal feature fusion module. In this way, we can make full use of the local and global feature information of the image. And a feature representation is generated, which can be more representative of the target category. The experimental results on the three public datasets of miniImageNet, tieredImageNet and CUB-200-2011 show that the proposed method can achieve better classification results. The accuracy rate of the proposed method on the 5way-5shot task reaches 81.69%, 85.36% and 89.78% respectively. Compared with the baseline model, the classification accuracy increased by 5.23%, 3.19% and 5.99% respectively.

Key words: image classification, few-shot learning, multi-scale features, attention mechanism, feature fusion