Cross Modal Few-Shot Contextual Transfer for Heterogenous Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning (FSL) in deep transfer learning struggles with limited data diversity, leading to biased local features.
• Existing methods often fail to capture reliable global features crucial for accurate classification in FSL scenarios.

Purpose of the Study:

• To propose a novel cross-modal few-shot contextual transfer method for enhanced few-shot image classification.
• To leverage contextual information from heterogeneous data sources to mitigate the limitations of sample specificity in FSL.

Main Methods:

• A cross-modal few-shot contextual transfer approach is developed, integrating textual and visual semantic information.
• Similarity measures are reformulated by fusing multi-modal information to inhibit sample specificity.
• A deep transfer scheme reuses pre-trained model extractors to enhance local feature extraction and recognition patterns.

Main Results:

• The proposed method effectively suppresses deviations caused by limited sample features in category definition.
• Integration of cross-modal and intra-modal contextual information significantly improves few-shot image classification accuracy.
• The approach demonstrates superior performance in scenarios with scarce training data.

Conclusions:

• Cross-modal contextual transfer is a viable strategy to enhance FSL by overcoming data scarcity and sample specificity.
• Fusing heterogeneous data sources provides robust feature representations for improved few-shot image classification.
• The method offers a promising direction for advancing deep transfer learning in low-data regimes.