Automated detection of damage precursors on specular surfaces using an enhanced deep-learning framework

Area of Science:

• Materials Science
• Computer Vision
• Optical Physics

Background:

• Industrial quality assurance for metal additive manufacturing (AM) faces challenges in detecting damage precursors on reflective surfaces.
• Conventional optical inspection methods are hindered by surface reflectivity, impacting small-defect recognition and environmental robustness.

Purpose of the Study:

• To propose an enhanced SOutlook-YOLOv11 framework for improved detection of damage precursors in metal AM components.
• To address limitations in small-defect recognition and environmental robustness for quality inspection.

Main Methods:

• Integration of polarized imaging with an optimized deep-learning architecture (SOutlook-YOLOv11).
• Introduction of the C3K2-Outlook module for enhanced feature representation.
• Insertion of a spatial adaptive feature modulation (SAFM) module for improved multi-scale feature fusion.
• Replacement of CIoU loss with Inner-CIoU loss for enhanced localization accuracy.

Main Results:

• The proposed model achieved a 4.2% increase in precision, 0.8% in recall, and 1.9% in mAP@50 compared to the baseline YOLOv11.
• A recall rate of 0.937 was reached, demonstrating high reliability in identifying damage precursors.
• Substantial performance improvements over the baseline YOLOv11 were observed.

Conclusions:

• The enhanced SOutlook-YOLOv11 framework offers a more reliable and intelligent solution for quality inspection in high-precision additive manufacturing.
• This work presents a novel strategy combining optical physics and deep learning for defect detection.
• The proposed method effectively overcomes challenges posed by reflective surfaces in metal AM quality assurance.