05 — CV · 2026
DriftSense
3.3×
better than classical NCC · MunichTech winner
NumPy · SciPy · OpenCV · FFT + LER fingerprinting
The problem
Classical template matching fails on the repetitive DRAM/FinFET lattices in SEM inspection images — every patch looks like every other patch.
The approach
FFT spectral analysis for candidate recall, then persistent line-edge-roughness (LER) fingerprinting for discrimination. No neural nets — pure signal processing.
The result
57.5% accuracy on the hard tier (±5° rotation, ±7% scale) vs 17.5% for classical NCC — 3.3× better; LER discrimination gave d′ ≈ 4–5 separability. 31 regression tests. MunichTech EXPO 2026 winner.
Code private / patent-pendingAll work