SEM image restoration
SemiRestore
Restoration is only useful when model output is paired with explicit provenance, diagnostics, and honest limits.
Open repositoryOverview
SemiRestore combines a statistics-conditioned NAF-SR restoration pipeline with a FastAPI service and an operational dashboard. It is designed around a narrow scientific imaging task rather than general photo enhancement.
Problem
Degraded SEM imagery can obscure fine structural detail. A restoration workflow must improve interpretability without presenting generated detail as measured ground truth.
Why it matters
Scientific-image restoration carries a higher burden of traceability than aesthetic enhancement. Diagnostics, model readiness, and input suitability need to remain visible alongside the output.
Architecture / Approach
- Checksum-gated model loading prevents silent fallback to an unverified checkpoint.
- A statistics-conditioned NAF-SR pipeline owns preprocessing, inference, and response provenance.
- The API remains live but reports unavailable readiness when the verified model artifact is absent.
- Uploads and restored images are processed in memory and are not permanently stored by the platform.
My contribution
- PyTorch model development and restoration-pipeline work.
- Inference, diagnostics, and validation workflows.
- Controlled evaluation using PSNR and SSIM without presenting scores as proof of correctness.
Limitations
Restored images are estimates, not ground truth. Diagnostics and suitability recommendations are advisory measurements and heuristics, not probabilities or proof of restoration correctness.
What I learned
Model quality and production reliability are inseparable: artifact verification, typed interfaces, bounded input handling, and transparent failure modes are part of the ML system itself.