Dynamic margin contrastive learning for open-set recognition in long-tailed sonar imagery.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Sonar image classification faces challenges with long-tail distributions and open-set recognition.
• Existing methods struggle to effectively handle imbalanced datasets and identify unknown classes.

Purpose of the Study:

• To introduce Dynamic Margin Contrastive Learning (DMCL), a novel framework for sonar image classification.
• To address challenges of long-tail distribution and open-set recognition simultaneously.
• To improve the robustness and accuracy of sonar image classification models.

Main Methods:

• DMCL utilizes a class-frequency-based dynamic margin mechanism for adaptive learning.
• A contrastive learning strategy is employed to generate robust feature representations.
• An uncertainty estimation module is incorporated for effective detection of unknown classes.

Main Results:

• DMCL demonstrated superior performance on the NKSID sonar image dataset compared to existing methods.
• Achieved significant improvements in Macro-F1 (89.47%), Normalized Accuracy (81.90%), OSCRmac (91.01%), and OSFM (93.21%).
• Outperformed the PLUD method by 5.79% in Macro-F1 and 5.87% in OSFM.

Conclusions:

• DMCL effectively handles long-tail distributions and open-set recognition in sonar image classification.
• The framework offers potential applications in domains with imbalanced data and unknown class scenarios.
• Validated effectiveness through comprehensive experimental results on a benchmark sonar dataset.