An Edge-Guided Feature Propagation Framework for Microdefect Inspection of Friction Stir Welds
Year: 2026
Type: article
Abstract: Defect detection in friction stir welding (FSW) is important for manufacturing safety, but remains difficult under degraded imaging conditions. Vibration blur weakens defect boundaries, repeated downsampling removes microdefect evidence, and complex weld textures introduce distracting edge responses. To address these issues, we propose an edge-guided adaptive microdefect-enhanced pyramid (EAMP) framework, instantiated mainly as EAMP-DEtection TRansformer (DETR), to preserve defect evidence during feature propagation. The framework contains three compact mechanisms. First, a microdefect adaptive feature pyramid uses the high-resolution$P_{2}$feature as auxiliary microdefect evidence and injects it into the$P_{3}$detection level, improving the representation of small defects without adding a high-resolution detection head. Second, an edge-guided representation extracts boundary cues from shallow features and propagates them into the semantic hierarchy to preserve weak contours. Third, modulation-based fusion adaptively combines boundary, semantic, and microdefect cues across scales to reduce interference from weld textures. The design is lightweight and plug-and-play, and can be instantiated beyond the DETR detector. Experiments on friction stir welding defect detection (FSW-DD) and public metallic defect datasets show that the proposed framework is effective and robust. On FSW-DD, EAMP-DETR improves mAP at 0.5 over real-time DEtection TRansformer (RT-DETR)-R18 by 5.42 percentage points with only a moderate parameter increase.
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
Award: G4246221281
OpenAlex ID: W7208739461