Collaborative Learning-based Dual Network for Few-Shot Image Classification

With the vigorous development of image classification technology in the field of computer vision, Few-shot learning (FSL) has become a research hotspot for solving classification task model training with a small number of samples. FSL aims to achieve efficient identification and processing of new category samples with few annotations. Previous works focus on information extraction based on one single model for FSL, lacking the distinction of the differences between data samples. Therefore, we present a meta-learning-based dual model with knowledge clustering for few-shot image classification, trying to learn the correlation between dual models and capture the information embedded in the data samples. In addition, we introduce the center loss to cluster the same sort of samples and to maximize the similarity among the intraclass and the difference among the inter-class. We adopt multiple tasks based on Meta-learning during the training stage. For each task, the training of dual models divides into two phases, which depend on each other under the guidance of the center loss. At the first phase, the first model is trained with a soft label obtained by the predicted label of the second model. The second phase repeats the information exchange of the first phase. We find that the optimal predictions of the active model are close to the soft and actual labels. Extensive experimental results on three general benchmarks illustrate the effectiveness of our proposed methods on few-shot classification tasks.