03 / Final degree project · 2024

NostalgIA

Restoring damaged historical photographs

A full experimentation pipeline that simulates grayscale, tears and scratches on CelebA images, compares GAN and U-Net approaches, and documents why U-Net became the final restoration architecture. Graded 10/10 at ETSISI, Universidad Politécnica de Madrid.

Python
U-Net / GAN
TensorFlow / Keras
OpenCV
CelebA

Context: Final degree project — ETSISI, UPM · June 2024
Grade: 10/10 · Advisor: Alberto Díaz Álvarez
Dataset: CelebA — ~210,000 face images
Code: Full pipeline on GitHub ↗

Restorations

3 visuals

Original, scratched and U-Net-restored portraits — Scratch restoration — original, damaged input, U-Net output

Original, torn and U-Net-restored portraits — Tear restoration — missing pieces rebuilt by the U-Net

Original photo, black-and-white input and GAN-recolored output — Colorization — the earlier GAN approach on black-and-white photos

How it works

Approach

Simulate the damage

Real damaged/undamaged photo pairs don't exist at scale, so grayscale, tears (Perlin noise, two thresholds) and scratches (Perlin + Canny) are synthesized on clean CelebA images.

Try the GAN first

A generator reconstructed damaged images against a discriminator. Promising on colorization, but it struggled with tears and scratches even after tuning.

Ship the U-Net

The damaged image goes in, the repaired image comes out. U-Net met the project goals and became the final architecture, documented end to end in the report.

See the proof
behind the case.

GitHub repo ↗LinkedIn profile ↗

← Back to selected work