Samsung patents method to train cameras to see hidden light details
Samsung describes a system that trains a computer model to convert standard color photos into detailed hyperspectral images. It uses a light meter to measure the exact lighting conditions and compares the model's output against known data to improve accuracy.
In the filing’s own words
An electronic device for training a spectral reconstruction model, including: at least one processor; and memory storing one or more instructions which cause the electronic device to: obtain a first red-green-blue (RGB) image of a scene including an object corresponding to a predetermined reflectance value, obtain a first illumination value corresponding to the first RGB image, provide the first RGB image and the first illumination value as input to the spectral reconstruction model to obtain a first hyperspectral image, obtain a first patch corresponding to the object from the first hyperspectral image, calculate a supervised loss based on a comparison between a first spectral value corresponding to the first patch and a ground truth spectral value determined based on the predetermined reflectance value and the first illumination, and train the spectral reconstruction model based on the supervised loss to obtain a trained spectral reconstruction model.
What exactly they claim (claim 1)
1. An electronic device for training a spectral reconstruction model, the electronic device comprising: a red-green-blue (RGB) camera configured to capture RGB images; a spectral photometer configured to measure illuminations to obtain illumination values; at least one processor; and memory storing one or more instructions which, when executed by the at least one processor, cause the electronic device to: obtain, using the RGB camera, a first red-green-blue (RGB) image of a scene comprising an object, wherein the object corresponds to a predetermined reflectance value, obtain a first illumination value corresponding to the first RGB image by measuring, using the spectral photometer, a first illumination of the scene when the first RGB image is captured, provide the first RGB image and the first illumination value as input to the spectral reconstruction model to obtain a first hyperspectral image, obtain a first patch corresponding to the object from the first hyperspectral image, calculate a supervised loss based on a comparison between a first spectral value corresponding to the first patch and a ground truth spectral value determined based on the predetermined reflectance value and the first illumination, and train the spectral reconstruction model by modifying at least one parameter of the spectral reconstruction model based on the supervised loss to obtain a trained spectral reconstruction model.
Facts
- Company
- Samsung (SAMSUNG ELECTRONICS CO., LTD.)
- Official title
- MIXED SUPERVISION FOR SPECTRAL RECONSTRUCTION
- Publication number
- US20260330828A1
- Application number
- 19640792
- Filed
- Apr 7, 2026
- Published
- Oct 22, 2026
- Inventors
- 5
- Claims
- 20
Classified under
G06V 20/194 Physics & computingG06N 3/09 Physics & computingG06T 7/90 Physics & computingG06V 10/60 Physics & computingG06T 2207/10024 Physics & computingG06T 2207/10036 Physics & computingMore from Samsung in Cameras & imaging
- Samsung patents syncing multiple cameras to record one video Sep 1, 2026
- Samsung patents camera system that auto-picks best lens for video Oct 13, 2026
- Samsung patents camera design that bends light to fit in thinner phones Oct 20, 2026
- Samsung patents camera feature that saves video clips around photos Sep 10, 2026
- Samsung patents image sensor that captures depth and bright scenes Sep 24, 2026
Summaries are simplified from the filing’s own abstract and first claim. A patent filing is not a product announcement.
Patent data and drawings: U.S. Patent and Trademark Office (USPTO). Nebula is not affiliated with or endorsed by the USPTO.
