Foster noisy label learning by exploiting noise-induced distortion in foreground localization

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks require large, well-annotated datasets for effective training.
• Noisy labels in datasets can significantly impair model generalization.
• Current methods for noisy label mitigation often rely on loss or confidence distributions.

Purpose of the Study:

• To investigate the impact of label noise on foreground localization using spatial attention.
• To propose a novel framework, FLSC, for robust noisy label learning.
• To enhance sample selection and label correction accuracy in the presence of noisy data.

Main Methods:

• Analyzing spatial attention distribution on visual activation maps to detect noise-induced distortions.
• Proposing a two-stage framework (FLSC) that integrates foreground localization with noisy label learning.
• Implementing a noise-adaptive adversarial erasing strategy to suppress background activation and improve representation learning.

Main Results:

• Noisy labels cause models to attend to irrelevant background regions or object edges.
• FLSC quantifies noise-induced distortion in foreground localization for better sample selection.
• FLSC demonstrates superior performance over state-of-the-art methods on synthetic and real-world datasets.

Conclusions:

• Evaluating localization quality based on feature activation is a novel approach to address label noise.
• FLSC effectively mitigates the negative effects of noisy labels on model training.
• The proposed method enhances the robustness and generalization of deep learning models.