PDeT: A Progressive Deformable Transformer for Photovoltaic Panel Defect Segmentation.

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Vision

Background:

• Defects in photovoltaic (PV) panels reduce power generation and can cause overheating.
• Precise pixel-level defect segmentation is essential for stable PV system operation.
• Existing methods struggle with adaptive scale determination and feature fusion for accurate defect localization.

Purpose of the Study:

• To propose a novel Progressive Deformable Transformer (PDeT) for precise defect segmentation in PV cells.
• To enhance feature extraction by adaptively determining receptive fields for accurate defect localization.
• To improve high-level representations through seamless fusion of semantic and fine-grained features.

Main Methods:

• Developed a Progressive Deformable Transformer (PDeT) incorporating adaptive spatial sampling offsets and self-attention.
• Implemented a semantic aggregation module for refining semantic information and balancing contextual information.
• Evaluated the PDeT on a dedicated solar cell dataset and the MVTec-AD dataset for cross-domain validation.

Main Results:

• Achieved a mean Intersection over Union (mIoU) of 88.41% on the solar cell defect segmentation dataset.
• Demonstrated superior performance compared to existing methods on PV cell defect detection.
• Showcased excellent recognition performance on the MVTec-AD dataset, validating cross-domain applicability.

Conclusions:

• The PDeT effectively addresses the challenges of adaptive scale determination and feature fusion in defect segmentation.
• The proposed method significantly improves the accuracy and robustness of defect detection in PV panels.
• PDeT shows promise for defect detection applications beyond PV cells, indicating its versatility.