Decision fusion for few-shot image classification

Recent few-shot learning methods mostly only use a single classifier to complete image classification. In general, a single classifier is likely to be overfitting because of its inherent drawbacks. However, the recognition rate of categorization will be significantly increased if we can utilize the complementary information of different classifiers. For the few-shot problem, the test samples come from new classes, which makes it difficult for a single classifier to distinguish, and it can be improved via decision fusion. In this paper, we propose decision fusion for few-shot learning (DF-FSL) to overcome the drawbacks of single classifier. To be specific, we assign the task to two classifiers, which are the logical regression classifier and probabilistic collaborative representation-based classifier (ProCRC), then allow the two classifiers to learn together through several iterations. Finally, we evaluate our approach on four benchmark image datasets, which include CIFAR-FS, CUB, miniImageNet and tieredImageNet datasets, and two remote sensing image datasets which are RSD46-WHU and NWPU-RESIS45. The experimental results illustrate the complementarity between different classifiers and show that the performance of our proposed DF-FSL method provides an obvious improvement. And DF-FSL can make great progress in few-shot remote sensing image classification.