Harnessing Group-Oriented Consistency Constraints for Semi-Supervised Semantic Segmentation in CdZnTe Semiconductors.

Area of Science:

• Materials Science
• Computer Vision
• Semiconductor Physics

Background:

• Labeling Cadmium Zinc Telluride (CdZnTe) semiconductor images presents challenges due to low-contrast defect boundaries, requiring cross-referencing multiple views.
• Existing semi-supervised semantic segmentation (SSS) methods are suboptimal for CdZnTe due to their inherent 'one-to-one' GT relationship, unlike CdZnTe's 'many-to-one' view-GT structure.
• This limitation can amplify errors in low-contrast areas and introduce confirmation bias during annotation.

Purpose of the Study:

• To develop a novel semi-supervised semantic segmentation framework tailored for the unique 'many-to-one' ground truth characteristic of CdZnTe semiconductor images.
• To address the limitations of traditional SSS methods in handling low-contrast defect boundaries and error accumulation.
• To propose a human-inspired, group-oriented approach for improved image labeling and defect detection in CdZnTe.

Main Methods:

• Introduced the Intra-group Consistency Augmentation Framework (ICAF), a group-oriented SSS pipeline.
• Validated group consistency constraints with Intra-group View Sampling (IVS) to establish a baseline.
• Developed the Pseudo-label Correction Network (PCN), comprising a View Augmentation Module (VAM) for boundary-aware view synthesis and a View Correction Module (VCM) for inter-view information interaction.

Main Results:

• The ICAF demonstrated significant effectiveness for CdZnTe materials, outperforming existing methods.
• Achieved a 70.6% mean Intersection over Union (mIoU) on the CdZnTe dataset using DeepLabV3+ with a ResNet-101 backbone.
• Required minimal group-annotated data (2 samples, 5‰), highlighting the framework's data efficiency.

Conclusions:

• The proposed ICAF effectively addresses the challenges of labeling CdZnTe semiconductor images by adopting a group-oriented perspective.
• The framework enhances segmentation accuracy, particularly in low-contrast regions, by leveraging intra-group consistency and advanced network modules.
• ICAF offers a promising solution for accurate defect boundary segmentation in CdZnTe with high data efficiency.