OpenVision — SemiCon-AI

AI-Powered Wafer Image Restoration for Semiconductor Inspection

Denoising · Super-Resolution · Reproducible ML Pipeline

---

Overview

Semiconductor wafer inspection systems introduce noise (Gaussian, Poisson), blur (Gaussian, motion), and low resolution that degrade defect-detection accuracy at nanometer scale.

OpenVision trains deep learning models to reverse these corruptions — producing cleaner, higher-resolution images without a paired real-world dataset by synthetically generating degradation from clean reference images.

Tasks

Task	Presets	Description
Denoising	light / medium / heavy	Remove Gaussian and Poisson noise, motion and Gaussian blur
Super-Resolution	x2 / x4	Upscale low-resolution wafer imagery

---

Project Structure

OpenVision/
├── configs/                        # Versioned YAML experiment configs
│   ├── degradation.yaml            # Base config (all tunable parameters)
│   ├── denoise_light.yaml          # Preset: mild noise, no blur
│   ├── denoise_medium.yaml         # Preset: balanced noise + occasional blur
│   ├── denoise_heavy.yaml          # Preset: aggressive noise + frequent blur
│   ├── sr_x2.yaml                  # Preset: super-resolution ×2
│   └── sr_x4.yaml                  # Preset: super-resolution ×4
│
├── datasets/                       # Core data pipeline
│   ├── __init__.py
│   ├── degradation.py              # Noise / blur / downsample + metadata schema
│   ├── wafer_dataset.py            # SEMPairDataset — crop, augment, degrade
│   └── logger.py                   # ExperimentLogger — appends to experiment.csv
│
├── scripts/                        # Utility scripts
│   ├── make_dummy_dataset.py       # Generate placeholder PNGs + manifest.json
│   ├── preview_pairs.py            # Visual QC grid + per-sample *_meta.json
│   └── validate_dataset.py        # Dataset health check (CI-friendly exit codes)
│
├── tests/                          # Unit test suite (37+ tests)
│   ├── __init__.py
│   ├── test_noise.py               # Gaussian + Poisson noise correctness
│   ├── test_downsample.py          # Downsample shape and range
│   ├── test_dataset.py             # Dataset loading, p=0/1, metadata schema
│   └── test_reproducibility.py    # Seed consistency and worker diversity
│
├── docs/                           # Documentation and planning
│   ├── degradation_pipeline.md     # Full technical reference
│   ├── SemiCon-AI_Hackathon_Execution_Plan.docx
│   ├── SemiCon-AI_Project_Plan_Summary.docx
│   ├── SemiCon-AI_Sprint_Planning.docx
│   └── team_weekly_plan.docx
│
├── data/                           # Clean reference images (gitignored)
├── outputs/                        # Generated previews and reports (gitignored)
├── checkpoints/                    # Model weights (gitignored)
│
├── .gitignore
├── requirements-degradation.txt
├── package.json
└── README.md

---

Quickstart

1. Clone and install dependencies

git clone https://github.com/SreeNaresh1/OpenVision.git
cd OpenVision
pip install -r requirements-degradation.txt

2. Generate dummy data

Skip this step once you have real wafer images — point --out at your clean image directory instead.

python scripts/make_dummy_dataset.py --out data/dummy_clean --n 20

3. Validate the dataset

python scripts/validate_dataset.py --data_dir data/dummy_clean
# Writes  →  outputs/validation_report.json
# Exits 1 →  if corrupt files or duplicates are detected (CI-safe)

4. Visually inspect degradation

python scripts/preview_pairs.py \
    --data_dir data/dummy_clean \
    --config   configs/denoise_medium.yaml \
    --num_samples 6
# Writes  →  outputs/degradation_preview.png
#         →  outputs/degradation_preview_meta.json

5. Run the full test suite

python -m pytest tests/ -v
# Expected: 37 tests, 0 failures

---

Configuration & Presets

All parameters live in versioned YAML files with a schema_version field. Never edit the base config directly — create a new preset file for each experiment.

Preset	Task	Description
denoise_light.yaml	Denoising	Mild noise, no blur
denoise_medium.yaml	Denoising	Balanced noise + occasional blur
denoise_heavy.yaml	Denoising	Aggressive noise + frequent blur
sr_x2.yaml	Super-Resolution	×2 upscaling
sr_x4.yaml	Super-Resolution	×4 upscaling

---

Tech Stack

Layer	Technology	Purpose
Language	Python 3.10+	Core implementation
Deep Learning	PyTorch 2.1+	Models, training, GPU, AMP
Image Processing	OpenCV 4.8+	I/O, blur, pixel degradation
Numerical	NumPy 1.24+	Noise generation, array math
Augmentation	Albumentations 2.0+	Geometric pair transforms
Metrics	scikit-image 0.22+	SSIM computation
Visualisation	Matplotlib 3.7+	Preview grids
Configuration	PyYAML 6.0+	Versioned experiment configs
Experiment Log	CSV (stdlib)	Zero-dependency run history
Testing	pytest	37+ unit tests
CI / CD	GitHub Actions	Lint + test gates on every PR

---

Experiment Tracking

Every training run is logged to outputs/experiment.csv automatically via ExperimentLogger:

timestamp | experiment_id | git_commit | config_name | epoch | loss | psnr | ssim

The git_commit column ties every result back to the exact code version — no external tracking tools required.

---

Degradation Metadata Schema

Every degraded sample carries a versioned metadata dict alongside the image tensor:

{
  "schema_version": "1.0.0",
  "sample_id": "dummy_003.png",
  "task": "denoise",
  "gaussian":   { "applied": true,  "sigma": 0.042 },
  "poisson":    { "applied": false, "peak": null },
  "blur":       { "applied": true,  "type": "gaussian", "kernel": 5 },
  "downsample": { "scale": null,    "interpolation": null }
}

---

Reproducibility

from datasets.wafer_dataset import SEMPairDataset, seed_worker
from torch.utils.data import DataLoader

train_ds = SEMPairDataset(
    root="data/train_clean",
    config="configs/denoise_medium.yaml",
    seed=42
)
train_loader = DataLoader(
    train_ds,
    batch_size=16,
    shuffle=True,
    num_workers=4,
    worker_init_fn=seed_worker   # ← required for deterministic training
)

Always pass worker_init_fn=seed_worker when num_workers > 0. Without it, every DataLoader worker inherits an identical RNG state, silently capping training diversity.

---

Hackathon Timeline

Phase	Dates	Deliverable
Phase 1 — Infrastructure	30 Jul – 01 Aug	Degradation pipeline, configs, 37 unit tests
Phase 2 — Dataset	02 Aug – 05 Aug	Validation pipeline, experiment logger, preview QC
Phase 3 — Baseline Training	06 Aug – 09 Aug	Denoising baseline, PSNR/SSIM benchmarks
Phase 4 — Evaluation	10 Aug – 12 Aug	SR ×2 inference, consolidated benchmark table
Phase 5 — Submission Prep	13 Aug – 15 Aug	Demo video, technical report, CI polish
Round 1 Deadline	16 Aug 2026	Final submission

---

Documentation

Document	Description
degradation_pipeline.md	Full technical reference for the degradation pipeline
SemiCon-AI_Project_Plan_Summary.docx	Project overview, architecture, and tech stack
SemiCon-AI_Sprint_Planning.docx	Master 10-week engineering roadmap
SemiCon-AI_Hackathon_Execution_Plan.docx	Compressed hackathon execution schedule
team_weekly_plan.docx	4-member weekly task and role-rotation plan

---

License

MIT License — see LICENSE for details.