Micro-cracks and residual stress in transparent components — architectural glazing, windshields, optical elements — are nearly invisible under ordinary lighting, yet they routinely precede failure.
PolarVision exploits stress birefringence. Placed between crossed polarizers, unstressed glass extinguishes the light and appears black, while stress concentrations rotate the polarization state and produce bright or colored interference fringes. The system captures these frames live from a USB camera, extracts crack geometry and stress distribution, and reports a four-level risk assessment with an interpretable score.
| Polarization-enhanced imaging | Brightness-weighted chroma response that isolates birefringence signal from sensor noise |
| 12-dimensional feature extraction | Crack length, width, area, orientation, curvature, branch and endpoint counts, tip gradient, texture entropy |
| Four-level risk grading | Intact / minor / medium / high, backed by an interpretable risk score R |
| Plug-and-play USB cameras | Automatic device enumeration, external-camera preference, hot-plug detection |
| Real-time processing | 15 fps capture with a frame-skipping pipeline; runs smoothly on low-spec laptops |
| Demo fallback | Cycles sample imagery when no camera is present, so the UI is never blank |
| One-click reporting | Composite PNG report plus a full JSON feature dump |
| Standalone executable | PyInstaller build requires no Python on the target machine |
[LED backlight] → [polarizer] → [specimen] → [analyzer ⊥] → [USB camera]
↓
unstressed → extinguished, appears black
stressed → birefringent, colored fringes appear
frame → polarization response S → CLAHE → validity mask → Canny edges
↓
risk grade ← risk score R ← 12-D features ← crack/stress masks ← morphology
Polarization response:
S = √( (V/255)² + Chroma² ) / √2 , Chroma = (Sat × V) / 255²
Saturation must be weighted by value. HSV saturation is unreliable at low brightness — a near-black pixel can report very high saturation — so using the raw channel turns sensor noise on a dark background into phantom stress signal.
Risk score:
R = w₁·(L/L_max) + w₂·(A_s/A_max) + w₃·(G_t/G_max) + w₄·(P_max/P_ref)
| R | Grade | Interpretation |
|---|---|---|
| < 0.15 | 🟢 INTACT | No significant defects |
| 0.15–0.35 | 🟡 MINOR | Minor cracking, monitor |
| 0.35–0.65 | 🟠 MEDIUM | Moderate risk, re-inspect |
| > 0.65 | 🔴 HIGH | High risk, replacement advised |
git clone https://github.com/<your-username>/PolarVision.git
cd PolarVision
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r requirements.txt
python main.py:: Keep the project in an ASCII-only path, e.g. C:\Projects\PolarVision\
build.batThe result lands in dist\PolarVision\PolarVision.exe. Ship the whole dist\PolarVision\ directory — the executable needs the _internal\ folder beside it.
python -m polarvision.cli path/to/image.jpg --out reports/PolarVision/
├── main.py entry point
├── build.bat / build.ps1 Windows build scripts
├── run_dev.bat run from source without packaging
├── build.spec PyInstaller configuration
├── polarvision/
│ ├── processing.py polarization response, enhancement, extraction
│ ├── features.py 12-D features, risk score, grading
│ ├── classifier.py rule-based grading + optional RandomForest
│ ├── cameras.py device enumeration and preference scoring
│ ├── welcome.py animated splash screen
│ ├── gui.py PyQt5 main window
│ ├── cli.py command-line interface
│ ├── report.py report generation
│ └── synthetic.py synthetic test image generator
├── sample_images/ photographed specimens + synthetic samples
└── docs/
├── PROJECT_BRIEF.md full technical documentation
└── images/
Six sliders adjust detection live. Optimal values depend on the optical path, camera, and ambient light.
| Slider | Controls | Raising it |
|---|---|---|
| Edge low / high | Canny hysteresis thresholds | Detects fewer faint edges |
| Saturation thr | Chroma floor for stress regions | Only more vivid fringes count as stress |
| Brightness thr | Value floor for stress regions | Suppresses false positives in low light |
| CLAHE clip | Contrast enhancement strength | Weak fringes stand out, noise amplifies too |
| Min crack area | Minimum connected-component size | Filters out more speckle noise |
Grading is rule-based by default. Once labelled specimens are available, a RandomForest can take over:
from polarvision.classifier import PolarVisionClassifier
# CSV header: L,W,A_c,theta,C,N_b,N_t,P_mean,P_max,A_s,G_t,H,label
PolarVisionClassifier.train_from_csv("samples.csv", out_path="model.joblib")Place the resulting model.joblib next to main.py (or the executable) and the application switches to ML inference at startup.
- The algorithm assumes input from a crossed-polarizer optical path. Frames captured under ordinary illumination have bright backgrounds throughout, which drives heavy false positives. This is an input-domain mismatch rather than an implementation defect.
- Single-angle capture cannot recover full Stokes parameters, so stress orientation is only coarsely estimated from image moments. A four-angle interface is present in
processing.pyfor a future hardware upgrade. RiskWeightsnormalization constants are empirical and should be calibrated against known reference specimens.- Ambient light contaminates stress detection; a darkroom or a light shield is recommended.
Python 3.12 · OpenCV · PyQt5 · NumPy · scikit-learn · scikit-image · PyInstaller · pygrabber
Project conception, system architecture, and hardware testing by Yichen Liang (Henan University of Science and Technology,CHINA).
Algorithm design, implementation, and software validation by Zihan Zhao (Penn State University,United States).
Fell free to ask if you have any questions through GitHub OR email: zihan_z64@outlook.com and yichenliang17@gmail.com.
Generative AI tools were used in parts of this project.


