Cross-Domain Few-Shot Learning Based on Data Corruption and Dual Attention Mechanism

Abstract In few-shot learning, the model must first be trained using the same type of dataset, and its effectiveness may be compromised when identifying classes from different datasets. Therefore, recent research has proposed cross-domain few-shot learning, which aims to train a single-network model that can be applied across most datasets. Cross-domain few-shot learning typically involves two stages during training: (1) pre-training and (2) meta-training. In the pre-training stage, a feature extractor is trained using traditional supervised learning to extract features. In the meta-training stage, a classifier is trained using few-shot learning methods, whereby meta-learning algorithms classify the features extracted by the feature extractor. In this study, a dual attention mechanism that incorporates channel and spatial attention is introduced into the feature extractor in the pre-training stage, to enhance feature extraction capabilities. In the meta-training stage, samples obtained from Fourier-transform-based image corruption algorithms are introduced. The transformation algorithm scales the phase and amplitude of the Fourier transform output to change the texture and semantics of the image, allowing the classifier to learn more diverse images and improve its generalization ability. Finally, during the meta-training process, the prediction losses of the images before and after corruption are simultaneously calculated to maintain consistency in the prediction results and effectively mitigating the risk of overfitting. The experimental results demonstrate that the proposed method achieves better performance on the majority of the eight testing datasets, effectively reducing overfitting during training.