ABC-Norm Regularization for Fine-Grained and Long-Tailed Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Real-world image classification faces challenges with complex data distributions, including fine-grained categories and long-tailed class imbalances.
• Existing methods often address fine-grained or long-tailed issues separately, lacking a unified approach.

Purpose of the Study:

• To propose a novel regularization technique, Adaptive Batch Confusion Norm (ABC-Norm), for simultaneously addressing fine-grained and long-tailed data distributions in image classification.
• To enhance model learning by introducing adversarial loss through adaptive classification confusion.

Main Methods:

• Constructing an adaptive batch prediction (ABP) matrix within each training batch.
• Developing the Adaptive Batch Confusion Norm (ABC-Norm) as a norm-based regularization loss.
• Coupling ABC-Norm with conventional cross-entropy loss to trigger adversarial learning.

Main Results:

• Demonstrated the efficacy of ABC-Norm on benchmark datasets representing real-world, fine-grained, and long-tailed scenarios (CUB-LT, iNaturalist2018, CUB, CAR, AIR, ImageNet-LT).
• Showcased ABC-Norm's ability to improve model learning effectiveness by introducing adaptive classification confusion.
• Validated the theoretical connection between ABC-Norm and rank minimization objectives.

Conclusions:

• ABC-Norm provides a simple, efficient, and unified solution for simultaneously handling fine-grained and long-tailed image classification problems.
• The proposed regularization technique strengthens model learning by leveraging adversarial principles.
• Experimental results confirm the superior performance of ABC-Norm compared to existing methods.