Self-supervised pairwise-sample resistance model for few-shot classification

The traditional supervised learning models rely on high-quality labeled samples heavily. In many fields, training the model on limited labeled samples will result in a weak generalization ability of the model. To address this problem, we propose a novel few-shot image classification method by self-supervised and metric learning, which contains two training steps: (1) Training the feature extractor and projection head with strong representational ability by self-supervised technology; (2) taking the trained feature extractor and projection head as the initialization meta-learning model, and fine-tuning the meta-learning model by the proposed loss functions. Specifically, we construct the pairwise-sample meta loss (ML) to consider the influence of each sample on the target sample in the feature space, and propose a novel regularization technique named resistance regularization based on pairwise-samples which is utilized as an auxiliary loss in the meta-learning model. The model performance is evaluated on the 5-way 1-shot and 5-way 5-shot classification tasks of mini-ImageNet and tired-ImageNet. The results demonstrate that the proposed method achieves the state-of-the-art performance.