CEA-DETR: A Multi-Scale Feature Fusion-Based Method for Wind Turbine Blade Surface Defect Detection.

Area of Science:

• Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Wind turbine blade surface defect detection is challenging due to scale variations, blurred textures, and complex backgrounds.
• Existing methods often suffer from insufficient accuracy and high false/missed detection rates.

Purpose of the Study:

• To propose an improved RTDETR-based detection framework, CEA-DETR, for efficient wind turbine blade surface defect inspection.
• To enhance the accuracy and efficiency of automated defect detection systems.

Main Methods:

• Designed a Cross-Scale Multi-Edge feature Extraction (CSME) backbone for fine-grained feature extraction.
• Constructed an Efficient Multi-Scale Feature Fusion Network (EMSFFN) for enhanced multi-scale defect representation.
• Introduced an adaptive sparse self-attention mechanism to improve focus on critical defect regions.

Main Results:

• CEA-DETR achieved mAP50 of 89.4% and mAP50:95 of 68.9%, outperforming the baseline by 3.1% and 6.5%.
• Reduced computational cost by 20.1% and parameter count by 8.1%.
• Demonstrated suitability for resource-constrained unmanned aerial vehicles (UAVs).

Conclusions:

• CEA-DETR offers a significant improvement in wind turbine blade surface defect detection accuracy and efficiency.
• The model's reduced computational footprint makes it ideal for real-time autonomous inspection using UAVs.