A Data Augmentation Method Based on Saliency Measures and Feature Transformation for Few-Shot Learning
Yanbiao Hao, Yihang Wei, Jianting Zhang · 2023
Few-shot learning is devoted to making a deep neural network recognize unseen categories with few labels. Because the number of labeled samples is limited, the actual data distribution is difficult to be reflected. As a result, the model has the problems of overfitting and poor generalization performance. To solve the issues above, a data augmentation method based on saliency measures and feature transformation is proposed for few-shot learning. In our method, the salient region measure module is utilized to select discriminative information and reduce redundant information, which makes contributions to improving the quality of generated pseudo data. Further, in order to make the augmented data help perfect the data distribution, a contextual feature transformation module is proposed to augment data in feature space. The results obtained by experimenting on two datasets indicate that the proposed data augmentation method performs better than other comparison methods in few-shot learning.