Task-Related Saliency for Few-Shot Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Metric-based few-shot classification methods are susceptible to distractions from task-unrelated objects or backgrounds.
• Limited support set samples in few-shot learning hinder accurate identification of task-related targets.

Purpose of the Study:

• To develop a novel method for metric-based few-shot learning that explicitly learns and utilizes task-related saliency features.
• To enhance the model's ability to focus on relevant features and ignore irrelevant information, mimicking human cognitive strategies.

Main Methods:

• Introduction of a Saliency Sensitive Module (SSM) trained with a multiclass classification task to enhance feature representation and locate saliency features.
• Development of a self-training-based Task-Related Saliency Network (TRSN) to distill task-related salience from SSM.
• A three-phase approach: modeling (SSM and TRSN), analyzing (TRSN extracts relevant features), and matching (discriminating samples by strengthening task-related features).

Main Results:

• The proposed method consistently improves performance across five-way 1-shot and 5-shot settings.
• Extensive experiments demonstrate significant performance gains on benchmark datasets.
• The method achieves state-of-the-art results in few-shot classification tasks.

Conclusions:

• Explicitly learning task-related saliency features significantly enhances metric-based few-shot classification.
• The proposed TRSN effectively extracts relevant features while suppressing irrelevant ones, leading to more accurate sample discrimination.
• The method offers a robust solution to the distraction problem in few-shot learning, achieving superior performance.