Five master's students from Politecnico di Milano and Politecnico di Torino have built a two-stage image-processing pipeline for beauty tech firm Dermaself that standardises lighting and exposure in selfies, then strips identifiable facial geometry through 3D reconstruction while re-projecting the original high-resolution skin texture.

The project, run under the Alta Scuola Politecnica joint honours programme, tackles two structural headaches for any brand deploying selfie-based skin analysis: inconsistent inputs and the regulatory and reputational exposure of storing identifiable faces. Dermaself's platform converts a selfie into a skin characteristic profile and matches the output against a retailer's own catalogue, so the quality and privacy of that first image directly determine both the recommendation engine's reliability and the brand's GDPR posture.

What does the pipeline actually do?

The system runs in two passes. Before capture, a guidance layer steers the user towards a usable distance and lighting. After capture, a hybrid illumination correction chain blends AI-driven correction with classical image-processing algorithms to flatten the differences that automatic exposure, white balance and on-device camera tuning introduce.

The anonymisation layer is the more technically distinctive element. Instead of blurring, masking or pixelating — operations that destroy the very features skin AI needs — the team reconstructs the face in three dimensions, modifies the underlying geometry to suppress biometric identifiers, then re-projects the original skin texture onto that modified surface.

The result, in the team's framing, separates two elements normally fused inside a single JPEG: identity and skin information. Spots, lesions, redness and texture survive; the person does not.

Who built it?

The five-strong team split across both Politecnico campuses:

• Andrea Germano and Adriano Giuliani, Computer Engineering, Politecnico di Torino
• Federico Greppi, Biomedical Engineering, Politecnico di Torino
• Edoardo Gribaldo, Computer Science and Engineering, Politecnico di Milano
• Alessia Soccionovo, Management Engineering, Politecnico di Milano

Academic supervision came from Professor Elisabetta Raguseo of Politecnico di Torino, who also serves as Director of Alta Scuola Politecnica, and Professor Federica Arrigoni of Politecnico di Milano. Dermaself acted as the industry partner, framing the brief around its existing AI skin analysis deployment rather than a purely academic use case.

Why this matters for formulators, retailers and compliance leads

For brands integrating selfie-based skin diagnostics in e-commerce, retail counters or event activations, the technical problem is reproducibility. The same consumer scanned under fluorescent retail lighting, warm home LEDs and an outdoor midday selfie can produce three different AI outputs for what is, biologically, the same face. A normalisation layer that flattens those variables reduces the noise floor on skin scoring and makes longitudinal self-tracking credible.

The privacy layer is the more strategic piece. Selfies are biometric data under GDPR and under most non-EU privacy regimes. Brands storing raw facial images carry an asymmetric risk: a breach exposes not just customer data but customer faces. A 3D-reconstructed, de-identified surface collapses that risk profile and also opens the door to dataset licensing for academic and clinical research, a recurring bottleneck in beauty AI development.

The combination is also commercially legible. Dermaself's model connects analysis output to a brand's or retailer's own product catalogue to generate personalised recommendations, so any improvement to input consistency directly tightens the conversion logic on the back end.

What is the next milestone?

The team describes early results as promising but stops short of clinical claims. The next gate is dermatologist-led validation of the anonymised images, confirming that the modified geometry does not introduce artefacts that skew downstream skin characteristic analysis. Brands evaluating Dermaself or comparable selfie AI stacks should watch for that validation read-out; it will determine whether the anonymised outputs can carry any dermatological weight in product matching, or remain confined to cosmetic recommendation use cases.

Watch also for dataset-licensing announcements. Privacy-preserving facial imaging is the rate-limiting step for training and benchmarking skin AI at scale, and a validated 3D anonymisation pipeline is one of the few credible technical routes to building shareable corpora without informed-consent bottlenecks.