Dual Feature Reconstruction Network For Few-shot Image Classification

Xiaowei Guo, Jijie Wu, Kai Ren, Qi Song, Xiaoxu Li

Abstract: Few-shot image classification aims to provide accurate predictions for novelty by learning from a limited number of samples. Classical few-shot image classification methods usually use data augmentation and self-supervision to compensate for the lack of training sample, and introduce migration learning and meta-learning to pre-train the model or accelerate the model optimization, which improves the classification performance of the model. However, with a small amount of labeled sample data, these methods cannot meet the requirements of the model’s ability to characterize sample features, resulting in a model that is highly susceptible to overfitting problems. In this paper, we propose a Dual Feature Reconstruction Network (DFRN) for few-shot image classification. The network constructs the double feature vector by two modules, in which the first-level feature module generates an attention mask based on the image to make the feature vector characterize more of the target region, and the secondary feature module interferes with the feature vector to improve its generalization performance. In addition, the network also enhances the classification performance of the model by considering the contextual information of the support classes through an auxiliary loss function. Through extensive experiments, the network proposed in this paper achieves excellent performance on Flowers, CUB and Cars datasets and outperforms other reference fine-grained image classification methods such as FRN.