A learning approach with incomplete pixel-level labels for deep neural networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Learning with incomplete labels in Neural Networks (NNs) is a growing research area.
• Existing methods primarily address global label incompleteness, overlooking local label deficiencies.
• Incomplete pixel-level labels present challenges due to missing local information.

Purpose of the Study:

• To develop a learning approach for Neural Networks that effectively handles incomplete pixel-level labels, specifically addressing local information loss.
• To propose a method that minimizes the impact of missing labels while leveraging available data.
• To validate the approach on diverse image segmentation tasks, including speech balloon extraction and medical imaging.

Main Methods:

• A novel learning approach utilizing two dynamic weighted maps (object and background pixels) in parallel.
• Integration of these dynamic maps into the loss function of target Neural Networks.
• Validation using speech balloon extraction from comic book images with algorithm-generated incomplete labels.
• Application to medical image segmentation to assess generalization.

Main Results:

• The proposed method achieved results comparable to state-of-the-art supervised approaches that use manual labels.
• Demonstrated effectiveness in speech balloon extraction without requiring any manual annotation.
• Successful application to medical image segmentation, confirming the approach's generalization capabilities.

Conclusions:

• The dynamic weighted map approach is a promising solution for learning with incomplete pixel-level labels in Neural Networks.
• The method significantly reduces the need for manual labeling in image segmentation tasks.
• The approach shows strong generalization across different domains, including creative and medical imaging.