A Multi-Modal Few-Shot Learning Framework for Foreign Object Segmentation in GIS Inspection.

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Gas-Insulated Switchgear (GIS) operation is vital for power system reliability.
• Automatic foreign object detection in GIS is difficult due to poor lighting and reflections.
• Existing methods struggle with precise segmentation in complex GIS environments.

Purpose of the Study:

• To develop a high-precision foreign object segmentation framework for GIS.
• To address challenges of low-light and reflective environments in GIS inspection.
• To enable rapid adaptation to new foreign object types with limited data.

Main Methods:

• Established a dual-light (visible and ultraviolet) image acquisition system.
• Designed a lightweight network for adaptive multi-modal feature fusion.
• Introduced a Multi-Similarity Guided Branch Network (MSBNet) for few-shot segmentation.
• Implemented background similarity guidance and multi-similarity fusion for enhanced segmentation accuracy.

Main Results:

• The proposed framework achieved high-precision foreign object segmentation in GIS.
• The dual-light system and fusion network effectively mitigated imaging interference.
• MSBNet demonstrated rapid adaptation to unseen foreign object classes with few samples.
• Significantly outperformed existing few-shot segmentation baselines on a GIS dataset.

Conclusions:

• The multi-modal few-shot learning framework offers a robust solution for GIS foreign object inspection.
• The approach enhances segmentation continuity and boundary accuracy.
• This method improves the safety and reliability of Gas-Insulated Switchgear operations.