Industrial vision
Anomaly detection for industrial inspection
Real inspection lines rarely hand you what a supervised model wants. Defects are rare, so the positive class is thin. Labels are coarse — an operator records that a part failed, not which pixels failed. And the camera, the lighting, and the product itself shift between one line and the next, so a model calibrated on a reference set degrades the moment it meets the floor.
- CDGP: Contrastive Dual Gaussian Processes for Weakly Supervised Anomaly Segmentation — AAAI 2027, under review, co-first author
- SPARC: Subspace Position-Aware Robust Few-Shot Calibration for Distribution-Shifted Industrial Anomaly Detection — AAAI 2027, under review, co-first author
- Wafer defect classification — in preparation for the International Journal of Production Research, first author
- A Satellite Image-Based Framework for Digitizing Port Yard States and Evaluating Spatial Operational Decisions — KSCM Spring Conference 2026, Best Paper Award
- Unsupervised anomaly detection — methodology under development, co-first author
- 3D Gaussian splatting — methodology under development, co-first author
Try it
The playground runs a deliberately small version of this idea in your browser, on real MVTec AD textures and WM-811K wafer maps — a memory bank of presumed-normal patches, and a nearest-neighbor search to find the one that matches nothing. You can race it.