Deep Attention-Based Imbalanced Image Classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Class imbalance in image classification leads to biased models favoring majority classes.
• This bias hinders learning discriminative features for minority classes, impacting overall performance.
• Existing methods often require prior knowledge of data distribution to adjust attention.

Purpose of the Study:

• To propose a novel attention-based approach, Deep Attention-based Imbalanced Image Classification (DAIIC).
• To automatically focus on minority classes in a data-driven manner without prior distribution knowledge.
• To enhance feature learning and classification performance on imbalanced image datasets.

Main Methods:

• Employed an attention network and a novel attention-augmented logistic regression function.
• Integrated attention assignment in both prediction and feature spaces for joint learning.
• Developed an objective function to automatically learn class-specific misclassification costs.

Main Results:

• DAIIC effectively encapsulates minority class features into discriminative learning.
• Learned misclassification costs guided attention networks for improved feature learning.
• Demonstrated superior performance over state-of-the-art methods on single-label and multi-label datasets.

Conclusions:

• DAIIC offers a generalizable and effective solution for imbalanced image classification.
• The method automatically learns to prioritize minority classes, mitigating bias.
• DAIIC shows strong potential for real-world applications with skewed data distributions.