Open-set long-tailed recognition via orthogonal prototype learning and false rejection correction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Learning from data with long-tailed and open-ended distributions presents significant challenges in machine learning.
• Accurate classification in such scenarios is crucial for real-world applications.

Purpose of the Study:

• To propose a novel dual-stream framework, OLPR (Open-set Long-tailed recognition based on orthogonal Prototype learning and false Rejection correction), to address the challenges of long-tailed and open-ended data distributions.
• To improve both closed-set classification accuracy and open-set recognition performance.

Main Methods:

• OLPR employs a dual-stream architecture with a Probabilistic Prediction Learning (PPL) branch for probability generation and a Distance Metric Learning (DML) branch for learning orthogonal class prototypes.
• The DML branch utilizes orthogonal prototype loss, balanced Softmin distance cross-entropy loss, and adversarial loss for compact open-set representation.
• An Iterative Clustering Module (ICM) is introduced to categorize open-set samples and correct false rejections by re-classifying misidentified known samples.

Main Results:

• Experiments on ImageNet-LT, Places-LT, CIFAR-10/100-LT, and a custom long-tailed open-ended dataset demonstrate OLPR's effectiveness.
• OLPR achieved up to a 2.2% improvement in overall classification accuracy in closed-set settings compared to state-of-the-art methods.
• In open-set settings, OLPR showed up to a 4% increase in F-measure.

Conclusions:

• The proposed OLPR framework effectively handles long-tailed and open-ended data distributions in image classification.
• OLPR's dual-stream approach with orthogonal prototype learning and false rejection correction offers significant performance gains in both closed-set and open-set recognition tasks.