AIRI and MIPT turned an ordinary RGB camera into a hyperspectral one — 31 values per pixel in a single exposure

📷 A regular RGB camera captures hyperspectral data — 31 values per pixel

A hyperspectral camera sees 31 values per pixel instead of three RGB channels — and all of it in a single exposure. A module from a group of researchers at AIRI, Samara University, and MIPT takes the place of a standard camera's lens and turns it into a hyperspectral one.

🔬 The difference from the classic approach lies in the optics and processing. Instead of a lens — a plate with a set of harmonic lenses, that is, diffractive optics. A neural network reconstructs the spectrum from the image: one exposure, then frame processing.

For comparison, existing hyperspectral cameras cost millions of rubles and capture a single frame over the course of a minute.

🍏 What the spectrum gives in practice: telling a real leaf from a plastic one, a ripe fruit from an unripe one, one pigment from another.

The paper “Diffract and Conquer: Hyperspectral Imaging from Any RGB Camera via Optical Encoding and Learning” has been accepted to IJCAI 2026, the presentation took place in Bremen — full text on Habr