WREN: Low Light Image Enhancement Using Retinex theory-based Double U-Net-like Structures

2026-07-17 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

WREN is a neural network designed for low light image enhancement (LLIE), leveraging Retinex theory to achieve robustness across diverse dynamic range scenes. Existing Retinex-based methods often struggle with unstable decomposition, leading to over-enhancement and sensitivity to varying lighting conditions. WREN addresses this with double U-Net-like structures, comprising two sub-networks. The first network, with one encoder and two decoders, decomposes an input image into reflectance and illumination maps. The second network, featuring a customized Transformer block, exclusively enhances the illumination map. This enhanced map is then recombined with the reflectance map. The network is trained end-to-end using a scale-invariant loss function, providing robustness against illumination scaling. Numerical results demonstrate WREN achieves "state-of-the-art" performance across multiple datasets.

Key takeaway

For Computer Vision Engineers developing low light image enhancement solutions, WREN offers a robust architectural blueprint. Its double U-Net-like structure, combined with a Transformer block for illumination map refinement, directly tackles the instability and dynamic range sensitivity common in Retinex-based methods. You should consider integrating similar decomposition and targeted enhancement strategies, alongside a scale-invariant loss, to improve your models' performance and stability across varied lighting conditions.

Key insights

WREN robustly enhances low-light images by using double U-Nets and a Transformer for stable Retinex decomposition and illumination refinement.

Principles

• Retinex theory decomposes images into reflectance and illumination maps.
• Stable decomposition prevents over-enhancement in low light image processing.
• Scale-invariant loss improves robustness against illumination scaling.

Method

WREN employs a first U-Net-like network for reflectance and illumination decomposition. A second U-Net-like network, with a Transformer block, then enhances only the illumination map, which is recombined. Training uses a scale-invariant loss.

In practice

• Apply WREN's architecture for robust low light image enhancement.
• Integrate Transformer blocks for targeted illumination map refinement.
• Utilize scale-invariant loss for illumination scaling robustness.

Topics

• Low Light Image Enhancement
• Retinex Theory
• U-Net
• Transformer Networks
• Image Decomposition
• Computer Vision

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Multinex: An ultra lightweight AI model advancing low light image enhancement — Multinex uses classical color vision and Retinex principles for efficient, compact low-light image enhancement.
• Continuous Splatting meets Retinex: Continuous Gaussian Splatting and Implicit Reflectance Modeling for Low-Light Image Enhancement — CGS-Retinex enhances low-light images by decoupling illumination and reflectance using continuous Gaussian splatting and implicit neural representation.
• TokenLight: Precise Lighting Control in Images using Attribute Tokens — Attribute tokens enable precise, continuous control over multiple illumination factors in image relighting.
• SWNet: A Cross-Spectral Network for Camouflaged Weed Detection — Cross-spectral fusion and edge refinement are crucial for accurate camouflaged weed detection in agriculture.
• PIXLRelight: Controllable Relighting via Intrinsic Conditioning — PIXLRelight enables fast, high-quality single-image relighting by integrating PBR with learned image synthesis via intrinsic conditioning.
• ATT-CR: Adaptive Triangular Transformer for Cloud Removal — Adaptive Triangular Transformer (ATT-CR) efficiently removes clouds from remote sensing images by combining linear triangular attention and feature-selective gating.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.