Abstract: The success of deep learning in image classification relies heavily on large-scale data.However,in many application scenarios,it is difficult to collect enough data for model training.Therefore,few-shot learning aimed at obtaining high-performance models with limited data becomes a hot research direction.In the field of few-shot image classification,using unlabeled data to augment the training datasets is a common method,but it faces two urgent problems:how to obtain pseudo-labels of unlabeled data and how to mitigate the negative impact of accumulated noise labels？ Firstly,in order to obtain high-quality pseudo-labels,it is necessary to solve the problem of noise labels caused by the distribution shift of the source domain and the target domain.A dependence enhancement method based on Hilbert-Schmidt independent criterion is proposed to improve the prediction reliability of pseudo-labels by maximizing the correlation between image feature representation and labels.Secondly,in order to overcome the problem of label prediction error that accumulates over time,a noise label interference reduction method is proposed to ensure that the gradient of samples with correct labels always dominates the training dynamics,so as to lead the model to the optimal solution.The above methods are evaluated on the benchmark datasets for few-shot image classification,namely mini-ImageNet and tiered-ImageNet.The results demonstrate that the proposed approach effectively utilizes unlabeled data to improve classification accuracy and achieves impressive classification performance.

Key words: Deep learning, Image classification, Few-shot learning, Pseudo-labels, Noise-labels, Hilbert-Schmidt independent criterion