Enhancing Semantic Segmentation through Reinforced Active Learning: Combating Dataset Imbalances and Bolstering Annotation Efficiency

This paper presents a novel framework called Reinforced Active Learning (RAL) to enhance semantic segmentation performance under limited annotation resources and class imbalance.

Key Contributions

• Reinforcement Learning-Based Sample Selection
  RAL employs a reinforcement learning agent to identify the most informative and diverse samples for annotation. The reward function is designed to improve segmentation performance while addressing class imbalance.

• Class Imbalance-Aware Sampling Strategy
  Introduces a metric that accounts for class distribution, encouraging the selection of samples that include rare or underrepresented classes. This leads to better performance across all categories.

• Annotation Efficiency
  RAL achieves high segmentation accuracy with fewer labeled samples, reducing the cost and time associated with manual annotation.

• Experimental Validation
  The framework was tested on datasets like Cityscapes and ACDC, where it outperformed conventional active learning methods in both efficiency and accuracy.

Conclusion

RAL effectively balances the need for performance and annotation efficiency, particularly in imbalanced datasets, making it a strong candidate for practical semantic segmentation applications.

Recommended citation: Han, D., Pham, H., & Cheng, S. (2023). "Enhancing Semantic Segmentation through Reinforced Active Learning: Combating Dataset Imbalances and Bolstering Annotation Efficiency." Journal of Electronic & Information Systems, 5(2), 45–60. https://doi.org/10.30564/jeis.v5i2.6063
Download Paper