Instance-Specific Semantic Augmentation for Long-Tailed Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Long-tailed classification methods often struggle with imbalanced data, leading to classifier overconfidence in majority classes.
• Existing data augmentation techniques lack instance-specific semantic transformations, limiting their effectiveness.

Purpose of the Study:

• To propose novel feature-level augmentation (FLA) and pixel-level augmentation (PLA) learning methods for long-tailed image classification.
• To address the overconfidence issue in head classes and improve the performance on tail classes.

Main Methods:

• A three-stage approach: learning feature space, modeling instance-specific semantic transformation ranges using Gaussian distributions and a semantic transformation generator (STG), and applying FLA/PLA for fine-tuning.
• STG is trained by constructing ground-truth distributions for head class instances.
• FLA generates feature augmentations, while PLA guides pixel-level augmentations.

Main Results:

• The proposed FLA and PLA methods significantly improve long-tailed image classification performance.
• The augmentation strategy is effective when combined with existing long-tail classification methods.
• Experiments on five imbalanced datasets demonstrate the method's effectiveness.

Conclusions:

• Instance-specific semantic transformations are crucial for effective data augmentation in long-tailed classification.
• The proposed FLA and PLA methods offer a flexible and powerful approach to enhance long-tailed image classification.
• This work provides a new direction for tackling data imbalance in deep learning models.