Few-shot Image Classification Method with Label Consistent and Inconsistent Self-supervised Learning

Image is an important form for storing and transmitting information. Deep neural network based supervised machine learning models have achieved great success in image classification task. However, the cost of collecting large amount of annotated data has limited the universality of deep learning in real-world applications. Recently, few-shot learning method addresses this challenge of learning with limited supervised data based on the idea of meta-learning. As a natural extension, previous works also try to integrate few-shot learning with self-supervised learning which can further improve the classification accuracy. These methods utilize the invariant features of samples, thus ignoring the latent diversity information and limiting the generalization performance. To tackle this problem, we adopt a feature fusion based framework to form pretext tasks for a better combination of self-supervised learning and few-shot learning. Specifically, in this paper, we firstly utilize the Mixup method to generate samples in the latent (embedding) space, and then use these samples to construct label consistent and inconsistent pretext tasks respectively. A more discriminative feature extractor is thus learned and results in a better classification performance. Moreover, we also use a self-attention module to extract the compact features to further improve the accuracy of the classification. Extensive experiments demonstrate the effectiveness of our proposed model on three popular benchmarks.