01 / THT
Read placement differences against a reference.
Polarity recognition with PCB-A-specific AI
REX uses PCB-A-specific AI to distinguish polarity marks and orientation cues. Comparing each component with a normal reference verifies its placement direction and detects reversed insertion.
Robust defect detection across component tilt and camera viewing angles
REX consistently detects defects by comparing the same component’s core features, even when its tilt or the camera viewing angle changes.
The same component, a different orientation
Prediction · Reversed
The polarity band appears on the opposite side of the capacitor. Check the orientation in the inputs, then inspect feature differences and attention separately.
Bright regions have fewer similar local features in the other image. Colour is not defect probability; imaging and surface variations can also cause a response.
How reference-based inspection works
- 01 / Reference comparison
Extract comparable features
The reference and inspected component pass through the same network.
- 02 / Spatial features
Distribute attention, retain position
Features from different component regions are compared together with their positions.
- 03 / Pair verification
Classify the relationship
Feature differences and similarities distinguish aligned, reversed, and wrong or missing parts.
A change in orientation is not always a defect. Component polarity and the applicable inspection criteria also matter.
THT VALIDATION
Performance, with evaluation context
Internal validation · 9,568 pairs Includes unseen boards
- Macro-F1
- 0.970
Equal-weight mean F1 across three classes
- Normal false-positive rate
- 0.65%
6,183 cross-board normal pairs
| Class | F1 | Pairs |
|---|---|---|
| Aligned | 0.993 | 7,796 |
| Reversed | 0.930 | 624 |
| Wrong / missing | 0.987 | 1,148 |
F1 is the harmonic mean of precision and recall. These are internal THT validation results, not measurements of the SMD crack model or a guarantee across production conditions.
