A team of five master's students from Alta Scuola Politecnica has built a two-part image-processing pipeline that standardises selfies for AI skin analysis and anonymises faces without destroying the skin detail the algorithms need. The work, developed directly with beauty-tech company Dermaself, targets the two biggest technical obstacles in selfie-based skin analysis: image inconsistency and identity exposure.

The consistency problem

A smartphone selfie that looks normal to a human eye can still misrepresent skin. Automatic exposure, white balance, on-device camera processing and ambient lighting all shift how skin appears from one capture to another. For an AI platform that must detect subtle visible characteristics — spots, redness, lesions, texture — these variations degrade both quality and comparability of results. The problem compounds when the same technology runs across a consumer's bathroom, a beauty retail floor, a brand event or an e-commerce flow.

The team's first solution addresses image normalisation at two points. Before capture, the system guides the user toward suitable conditions, including correct distance and lighting. After capture, a hybrid pipeline combines AI-driven illumination correction with classical algorithms to remove residual variation. The goal, stated plainly: give Dermaself's AI consistent inputs regardless of where or how the selfie was taken.

The privacy problem

A facial image carries two types of information simultaneously — skin characteristics and identity. Traditional anonymisation techniques such as blurring, masking or pixelating a face protect identity but strip out precisely the skin details required for analysis and dermatological research.

Rather than blurring, the students built a 3D-based facial reconstruction and anonymisation pipeline. The system reconstructs the face in three dimensions, modifies its underlying geometry to reduce biometric identifiers, then re-projects the original high-resolution skin texture onto the altered structure. Visible skin information — spots, lesions, redness, texture — is preserved while the person in the image becomes unrecognisable. In effect, the technology separates two elements that a selfie normally fuses: identity and skin information.

The team

Alta Scuola Politecnica, the joint honours programme of Politecnico di Milano and Politecnico di Torino, supplied a deliberately multidisciplinary roster. Politecnico di Torino contributed Andrea Germano and Adriano Giuliani (Computer Engineering) and Federico Greppi (Biomedical Engineering). Politecnico di Milano contributed Edoardo Gribaldo (Computer Science and Engineering) and Alessia Soccionovo (Management Engineering). Professor Elisabetta Raguseo of Politecnico di Torino, who also serves as Director of Alta Scuola Politecnica, and Professor Federica Arrigoni of Politecnico di Milano supervised the project.

Integration with Dermaself

Dermaself's platform lets beauty brands and retailers embed selfie-based skin analysis into digital and physical customer experiences. From a single selfie, the technology analyses visible skin characteristics and maps the results against the brand's own product and ingredient catalogue to generate personalised skincare recommendations — across e-commerce platforms, retail environments and events.

The new technologies slot into that ecosystem. Image standardisation makes selfies captured in different settings more comparable; anonymisation offers a route to protecting the identities of the people in those images. Early results from the research are promising, and the team frames broader validation — including evaluation by dermatologists — as the project's next stage.

Why it extends beyond retail

The privacy architecture has implications well beyond product recommendations. Beauty companies, research organisations and clinics increasingly need high-quality image datasets to develop and validate AI skin analysis and digital dermatology tools, yet collecting and sharing facial images raises substantial privacy problems. A system that preserves relevant skin characteristics while reducing identifiable biometric information could make it easier to build, store and potentially license research datasets while better protecting the individuals represented in them.

The project ultimately poses — and offers one technical answer to — a question now facing the wider beauty industry: can AI skin analysis grow more accurate without compromising consumer privacy? With dermatologist review still pending, the collaboration at minimum demonstrates a working model for translating university research into deployable beauty-tech infrastructure.