A novel competitive learning prototype for image ordinal classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Image ordinal classification (IOC) assigns ordered labels to images, like age estimation.
• Metric learning is common for IOC, but incorporating ordinal information is challenging.
• Existing methods often overlook the direct use of ordinal properties in classification.

Purpose of the Study:

• To propose a novel Competitive Learning (CL) prototype for image ordinal classification.
• To effectively integrate ordinal information into the metric learning framework for IOC.
• To enhance feature discriminative power and improve the accuracy of ordinal classification tasks.

Main Methods:

• Adopted mixing techniques (Mosaic, CutMix, Mixup) from data augmentation for metric learning.
• Embedded ordinal measurements of image pairs via virtual combinations for enhanced feature learning.
• Utilized the difference between ordinal values to measure subtle distinctions in IOC.
• Introduced dual and random augmentation strategies to increase feature robustness.

Main Results:

• The proposed Competitive Learning (CL) approach effectively incorporates ordinal information.
• The method demonstrated enhanced discriminative power for local features.
• Experiments on age and car date estimation showed significant performance improvements over prior methods.
• The approach proved effective and robust across different ordinal classification tasks.

Conclusions:

• Competitive Learning (CL) offers a powerful new paradigm for image ordinal classification.
• Adapting mixing techniques for metric learning provides a simple yet effective solution.
• The method successfully addresses the challenge of integrating ordinal information into classification.