Image super-resolution process

FR3126526B1Active Publication Date: 2024-11-08SAFRAN ELECTRONICS & DEFENSE (FR) +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
FR2021009177
Authority / Receiving Office
FR · FR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2021-09-02
Publication Date
2024-11-08
Estimated Expiration
2041-09-02

AI Technical Summary

Technical Problem

Existing recurrent neural network methods for image super-resolution in quasi-static scenes suffer from stability issues and artifacts, while maintaining high quality and reducing computational cost remains a challenge.

Method used

A supervised training method for an artificial neural network with specific constraints on convolution layers, including a strict Lipschitz constraint on the second part, minimizes artifacts by ensuring stability and coherence in super-resolved images.

Benefits of technology

The method produces high-quality, artifact-free super-resolved images without increasing computational cost, maintaining temporal coherence and stability in quasi-static scenes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000016_0000
    Figure 00000016_0000
  • Figure 00000016_0001
    Figure 00000016_0001
Patent Text Reader

Abstract

Image Super-Resolution Method. One aspect of the invention relates to a method for obtaining a sequence of super-resolved images from a sequence of ultra-high-resolution images to be super-resolved, said method comprising the following steps: Supervised training of an artificial neural network on a database to enable it to provide a super-resolved image from a subsequence of ultra-high-resolution images, said database comprising a plurality of ultra-high-resolution image sequences, each associated with a sequence of super-resolved images, said neural network comprising a first part, a second recurrent part, and a third part, each comprising a set of convolutional layers, each corresponding to a circulating matrix, the supervised training step comprising, for each sequence of ultra-high-resolution images in the database,A substep of updating artificial neural network parameters by minimizing a cost function corresponding to the error between a super-resolved image sequence provided by the artificial neural network from the super-resolved image sequence, and the super-resolved image sequence associated with the super-resolved image sequence in the database, the minimization being achieved by applying a constraint on the spectral norm value of each circulating matrix corresponding to a convolution layer of the second part; Use on the super-resolved image sequence to be super-solved, of the artificial neural network trained to obtain a super-resolved image sequence.
Need to check novelty before this filing date? Find Prior Art