Aesthetica AI: Automated Image Management for Aesthetic Surgery

This project represents my first freelance delivery as an independent developer after leaving my full-time role to work remotely and on my own terms.

I developed a fully local, AI-based system designed to automate the management of pre- and post-operative images for aesthetic surgery.
The objective was to ensure visual consistency, accurate comparisons, and structured organization of patient photos using a combination of computer vision and image processing techniques.

The Project

Phase 1 – Recognition and Matching

The primary goal was to automate what is traditionally a fully manual and time-consuming workflow.

• Automatic pose classification: Front, back, and profile detection with ~95% accuracy.
• Intelligent pre/post matching: Reliable pairing even when multiple images exist for the same pose.
• Automatic generation: Creation of standardized side-by-side comparison images.
• Structured storage: Export in .webp format with consistent naming conventions and per-patient folder structures.

Phase 2 – Technical Optimization and Image Normalization

To achieve professional and unbiased results, the system applies several automatic refinements:

• Adaptive zooming: Detection of anatomical reference points (e.g. hips) to preserve consistent proportions.
• Automatic subject segmentation: Background removal with replacement options (solid color, gradient, or custom image).
• Visual normalization: Automatic contrast, brightness, and color adjustments to align photos taken under different conditions.
• Geometric alignment: Improved left/right profile matching through classical geometric techniques combined with AI predictions.

Technical Notes

The system deliberately avoids arbitrary top/bottom cropping to prevent framing errors and comparison bias.
Instead, proportions are preserved through visual compensation and alignment strategies.

This project clearly highlighted some current limitations of AI-based vision models.
In particular, distinguishing left vs right profile images in real-world, unconstrained datasets proved unreliable when relying on AI alone. The final solution required combining model predictions with classical statistical and geometric methods, rather than trusting a single model output.

The pipeline is probabilistic by nature (as with any AI system), but after several days of iteration and testing, it proved to be stable, accurate, and robust on real clinical datasets.

Conclusion

This project took more iterations than initially expected, but pushing through edge cases is what ultimately made the system solid.

It reminded me why I enjoy freelance work:
solving real-world problems, understanding where AI falls short, and engineering practical solutions beyond the hype.

This was a technically demanding challenge, and I deliberately invested extra time to ensure the final result fully met professional expectations.
Looking back, I’m confident this was one of the most complex and rewarding engineering problems I’ve tackled so far.