Tell me about a challenging project and how you solved it

"I’d like to share the FrameInspect-AOI project—Samsung smartphone mid-frame anodizing defect detection system, our largest vision inspection project with 150 million RMB revenue and 30+ production lines.

The Challenge: Samsung required detecting micro-scratches and anodizing color differences on mid-frame surfaces. Defect sizes were as small as 0.05mm, with <50ms inspection time and <0.3% false positive rate. These defects had extremely low contrast—some only visible under specific lighting angles.

Technical Evolution: I started with traditional vision algorithms—dynamic thresholding for initial screening, morphological operations for defect region extraction, and GLCM (Gray Level Co-occurrence Matrix) for texture classification. This worked for obvious defects but failed on low-contrast micro-scratches.

Key Innovation: I led the transition to deep learning with a CNN-Transformer hybrid architecture:

• YOLOv8-P2 for real-time preliminary screening of suspicious regions
• Swin Transformer for fine segmentation with global context understanding
• Traditional Blob Analysis for final confirmation, ensuring interpretability
• This cascade architecture reduced false positives by 35% and improved detection rate by 18%

My Role: As Technical Project Manager, I led an 8-person algorithm team, designed the multi-spectral optical solution (coaxial + low-angle ring + UV auxiliary light), and managed the full lifecycle from DOE validation to mass production deployment. The project passed Samsung’s stringent MSA (Gage R&R) validation and achieved zero-defect mass production."

1. “Traditional algorithms are the foundation, deep learning is the building—industrial scenarios need both.”
2. “Optical design solves 70% of problems; algorithms only handle the remaining 30%.”
3. “Zero-defect mass production isn’t achieved by algorithms alone—it requires system-level validation: optics + algorithm + mechanics + process.”