Low-illumination image spectral reconstruction and color restoration method based on frequency domain cooperative driving

By employing a frequency-domain collaborative-driven method for spectral reconstruction of low-light images, and utilizing Laplacian pyramid decomposition and dual-domain feature co-evolutionary units, combined with gating interaction mechanisms and residual accumulation, the color shift problem in spectral reconstruction under low-light conditions is solved, achieving accurate reconstruction and natural color restoration of hyperspectral images.

CN122115286APending Publication Date: 2026-05-29YUNNAN NORMAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
YUNNAN NORMAL UNIV
Filing Date
2026-01-28
Publication Date
2026-05-29

AI Technical Summary

Technical Problem

Under low-light conditions, existing technologies struggle to effectively distinguish noise from high-frequency texture details in low-light images, leading to color shifts and feature aliasing during spectral reconstruction, which affects the color fidelity and naturalness of the image.

Method used

A frequency-domain collaborative spectral reconstruction method for low-light images is adopted. By constructing a frequency-domain collaborative spectral reconstruction network, utilizing Laplacian pyramid decomposition and dual-domain feature co-evolutionary units, combined with gating interaction mechanisms and residual accumulation, the low-frequency structure flow and high-frequency texture flow are gradually optimized to reconstruct hyperspectral images. Furthermore, a spectral visual mapping module based on content-adaptive spectral weighting and neural color rendering network is constructed to achieve local adaptive tone mapping.

Benefits of technology

It significantly improves the robustness and generalization ability of image reconstruction under low illumination conditions, generates hyperspectral images with accurate colors and natural light and shadow levels that conform to human visual perception, and solves the problem of color shift under low illumination.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122115286A_ABST
    Figure CN122115286A_ABST
Patent Text Reader

Abstract

The application obtains a data pair of a low-illumination image and a hyperspectral image without motion blur at the same time and the same angle in a real scene by acquiring an RGB image and a hyperspectral image data pair, combining a Poisson-Gaussian mixed noise model and an exposure decay model, constructing a progressive frequency domain collaborative spectral reconstruction network, decoupling the image into a low-frequency structure stream and a high-frequency texture stream by using the Laplacian pyramid principle, designing a dual-domain feature collaborative evolution unit, using the core component MASD to perform parallel optimization on the spectral consistency of the structure stream and the spatial details of the texture stream, and combining a gating interaction mechanism and residual accumulation to gradually refine the image from blur to clarity, so as to reconstruct an accurate spectrum. A spectral-visual mapping module based on content adaptive spectral weighting and neural color rendering network is further constructed to replace the traditional fixed chroma matching function and map the reconstructed spectral image into an enhanced RGB image conforming to the human eye visual characteristics.
Need to check novelty before this filing date? Find Prior Art