Category Alignment Mechanism for Few-Shot Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Existing metric-based few-shot classification methods struggle with novel tasks due to feature embeddings that fail to encode discriminative properties effectively.
• Current matching methods inadequately utilize support set samples, neglecting contrastive relationships across categories for discriminative features.

Purpose of the Study:

• To develop an adaptable few-shot image classification method that enhances feature embeddings for novel tasks.
• To improve the utilization of support set samples by exploiting intra- and inter-category contrastive relationships.

Main Methods:

• Proposed a Category Alignment Mechanism (CAM) to align query image features with different categories, ensuring distinctness and strong correlation to contrastive relationships.
• Implemented a parameter-free, training-free CAM that adjusts features for matching when task categories change.
• Utilized a cross-validation-based feature selection for support samples to generate more discriminative category prototypes.

Main Results:

• The proposed method demonstrates consistent performance improvements on benchmark few-shot image classification tasks.
• The algorithm surpasses current state-of-the-art methods in both inductive and transductive inference settings.
• Extensive experiments on six datasets validate the effectiveness of the Category Alignment Mechanism.

Conclusions:

• The Category Alignment Mechanism (CAM) effectively makes feature embeddings adaptable and category-related for few-shot image classification.
• The method significantly enhances discriminative feature extraction by leveraging contrastive relationships, leading to superior performance on novel tasks.