Screen Panel Anomaly

Automated detection of subtle surface anomalies on industrial display panels using structured illumination and deep learning.

Problem

Surface defects on display panels are often extremely small, low-contrast, and visually similar to background noise, making manual inspection unreliable and time-consuming.
The goal of this project was to develop an automated pipeline capable of detecting and localizing microscopic panel defects under controlled illumination conditions.

Key constraints included:

• Defects occupying only a few pixels
• Strong background patterns (moving stripes, lighting artifacts)
• Severe class imbalance between defective and non-defective regions
• Requirement for high precision to avoid false alarms in production

Approach

Data

• Custom image acquisition setup using:
  • Fixed camera
  • Display monitor with structured illumination (white light + moving stripes)
  • Controlled background lighting
• Multiple captures of the same panel were merged to enhance defect visibility
• Two defect classes annotated using bounding boxes
• Dataset characterized by extreme foreground sparsity

Model

• YOLO-based object detection framework for real-time defect localization
• Two-class detection setup
• SAM explored for post-processing and region refinement
• Detection focused on small-object sensitivity rather than large contextual cues

Training

• Long-horizon training (≈400 epochs) to stabilize learning under class imbalance
• Class-wise loss monitoring to ensure balanced convergence
• Extensive filtering of false positives during early epochs
• Careful threshold tuning for confidence-based suppression

Evaluation

• Precision and recall tracked independently for each defect class
• Qualitative inspection used as a primary validation signal due to rarity of defects
• Final models achieved stable high precision and recall after convergence

Post-evaluation detection results on unseen panels. The model localizes subtle defects with high confidence while suppressing illumination-induced artifacts.

Notes & Lessons Learned

• Structured illumination is critical: raw images without stripe modulation hide most defects
• Merging multiple frames significantly improves defect visibility
• Small-object detection is more sensitive to preprocessing than model choice
• Early training instability is expected due to extreme imbalance and sparse positives
• Classical image processing remains valuable as a complement to deep learning in industrial settings

This project emphasized robust deployment-oriented design, balancing detection accuracy with false-positive control under real manufacturing constraints.