Enhancing Few-Shot Image Classification With a Multi-Faceted Self-Supervised and Contrastive Learning Approach

One effective approach for solving few-shot classification is learning deep representations that measure the similarity between query images and a few support images of specific categories. Recent methods overly relied on a single metric, resulting in insufficient interdependence in image feature representations. While more powerful metric methods can address this issue, they inevitably lead to severe overfitting. Balancing this contradiction is crucial in few-shot learning. Therefore, this paper proposes an enhanced feature reconstruction network based on holistic self-supervised tasks, learning features from multiple perspectives to balance this issue. Firstly, by combining feature reconstruction with distribution modeling, we enhance the metric algorithm using Brownian distance covariance constraint for feature reconstruction. Secondly, we employ a self-supervised training strategy for the overall task. Unlike previous upstream-downstream separated self-supervised frameworks, we use self-supervised transformations to enhance samples in the overall task, combined with contrastive learning. This strategy enables the model to perceive different transformations and feature invariances effectively, alleviating overfitting issues during metric method improvements. Experimental results on multiple datasets demonstrate that our method effectively improves the accuracy of few-shot image classification and achieves competitive results in cross-domain tasks and inference speed. Also, we utilized the feature extraction capability of the DINOv2 baseline self-supervised learning model, proving that our projection into self-supervision approach has strong adaptability and feasibility.