基于伪标签依赖增强与噪声干扰消减的小样本图像分类

Few-shot Image Classification Based on Pseudo-label Dependence Enhancement and Noise Interference Reduction

深度学习在图像分类领域的成功很大程度上依赖于大规模数据,然而在许多应用场景中,收集足够的数据用于模型的训练是比较困难的。因此,旨在利用有限的数据获得高性能模型的小样本学习成为热点研究方向。在小样本图像分类领域,使用无标签数据来扩充训练数据集是一种常用的方法,但该方法面临两个亟待解决的难题:如何获取无标签数据的伪标签以及如何减轻噪声标签累积的负面影响?首先,为获得高质量的伪标签,需要解决由源域和目标域的分布偏移导致的噪声标签问题,因而提出基于希尔伯特-施密特独立准则(Hilbert-Schmidt Independent Criterion,HSIC)的依赖增强方法,通过最大化图像特征表示与标签之间的相关性,从而提高伪标签的预测可靠度。其次,为克服标签预测误差随着时间推移不断累积的问题,提出噪声标签干扰消减(Noise Label Interference Reduction,NLIR)方法,确保具有正确标签的样本的梯度始终主导着训练动态,从而将模型引向最优解。所提方法在小样本图像分类基准数据集mini-ImageNet和tiered-ImageNet上进行了评估,实验结果表明,该方法能够很好地利用无标签数据提升分类精度,具有良好的分类性能。

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.