retinex decomposition

This work addresses the challenge of accurately decoupling reflectance and illumination components in low-light image enhancement. To this end, the authors propose a two-stage Retinex-guided Transformer model. In the first stage, the multiplicative Retinex model is transformed into an additive form in latent space through a logarithmic transformation with a one-pixel offset, enabling stable decomposition. The second stage employs a U-shaped refinement module that integrates Transformer-based attention mechanisms to enhance detail recovery and refine illumination distribution. By innovatively combining latent-space decomposition with a guided Transformer architecture, the method achieves competitive enhancement performance across four benchmark datasets while significantly improving training stability.

A key challenge in Retinex-based low-light enhancement is the incomplete decoupling of illumination and reflectance components, leading to residual inter-component residuals (ICRs) that cause perceptual distortion. To address this, we propose IRetinex—a co-optimized decomposition-and-enhancement framework featuring mutual correction. For the first time, IRetinex explicitly models and suppresses ICRs via a cross-component residual suppression module and a feature-similarity-guided residual correction mechanism, enabling joint illumination-reflectance disentanglement and fidelity-preserving enhancement within a deep learning framework. Extensive experiments on three benchmark low-light datasets—LOL, MIT-Adobe FiveK, and SID—demonstrate that IRetinex consistently outperforms state-of-the-art methods. Quantitative evaluations show significant improvements across PSNR, SSIM, and LPIPS metrics, while qualitative results confirm enhanced visual quality and superior detail restoration.

Low-light images commonly suffer from color shifts, low contrast, and noise corruption, degrading performance in downstream vision tasks. To address four fundamental limitations of classical Retinex models—channel-wise independence, lack of neuroscientific grounding, irreversible reconstruction, and inability to explain color constancy—this paper proposes the first quaternion-based Retinex model. It employs quaternions to jointly represent RGB channels and models reflectance–illumination coupling via Hamiltonian multiplication. A novel reflectance consistency metric is introduced to quantify color constancy. Furthermore, a dedicated quaternion neural network enables end-to-end invertible decomposition. Evaluated on low-light crack detection, multi-illumination face detection, and infrared–visible image fusion, the method outperforms state-of-the-art approaches by 2–11%, achieving significant improvements in color fidelity, noise suppression, and reflectance stability.

This work addresses the challenges of decomposition instability and over-enhancement in low-light image enhancement under complex dynamic range scenarios. Building upon Retinex theory, the authors propose a dual-branch U-Net-like architecture: the first branch decomposes the input image into reflectance and illumination components, while the second branch enhances only the illumination map. A customized Transformer module is introduced to effectively model long-range illumination dependencies. The method rigorously adheres to Retinex assumptions and incorporates a scale-invariant loss function to improve robustness against illumination scaling. Extensive experiments demonstrate that the proposed model achieves state-of-the-art performance across multiple benchmark datasets, significantly enhancing the stability and generalization capability of low-light image restoration.

This work addresses the challenge of low-light image enhancement, where balancing global illumination adjustment and local high-frequency detail recovery often leads to color distortion and structural artifacts. The paper introduces continuous Gaussian splatting into this task for the first time, proposing a novel continuous physical representation paradigm grounded in Retinex theory. By jointly modeling illumination and reflectance through an explicit–implicit framework, the method employs a continuous Gaussian renderer to estimate spatially smooth global illumination while representing reflectance via an implicit neural function, guided by high-frequency features for texture reconstruction. Brightness consistency constraints and illumination smoothness regularization are incorporated to eliminate grid-like artifacts caused by discrete sampling, enabling precise decoupling of illumination and texture. This approach effectively suppresses noise and overexposure while significantly improving high-frequency structural fidelity and color accuracy, outperforming existing methods.

To address illumination non-uniformity, noise corruption, and brightness-color imbalance in low-light images, this paper proposes a joint enhancement framework integrating Retinex decomposition with a Conditional Rectification Flow (CRF). Methodologically, the approach first decomposes an image into illumination and reflectance components via a Retinex model; second, it employs CRF to explicitly model the continuous spatial variation of illumination for adaptive lighting adjustment; third, it introduces a flow-field-driven data augmentation strategy to jointly train a deep denoising network, enabling simultaneous noise suppression in the reflectance domain and customizable brightness enhancement. Extensive experiments demonstrate that the method achieves state-of-the-art performance on multiple low-light enhancement and exposure correction benchmarks, yielding consistent improvements in quantitative metrics (PSNR/SSIM) and visual quality. The framework exhibits strong robustness to diverse degradation patterns and maintains high practicality for real-world deployment.

This work addresses the challenge of segmenting visually entangled objects—such as camouflaged, transparent, or defective targets—in RGB images. Inspired by Retinex theory, the authors decompose an image into illumination and reflectance components to enhance foreground-background discriminability within a unified spatial domain. They introduce the “Discriminability Gap Theorem,” which theoretically guarantees that this decomposition preserves or even improves discriminative capacity across diverse concealed scenarios. Building upon this insight, they design a task-driven Retinex decomposition module, a discriminability gap attention mechanism, and a camouflage-breaking contrastive loss operating in the reflectance feature space. Extensive experiments on multiple concealed object segmentation subtasks demonstrate significant performance gains, validating the effectiveness and generalizability of their homogeneous decomposition strategy.

This work addresses the challenges of deploying low-light image enhancement models—namely, excessive model size, reliance on a single color space, and resulting color or exposure artifacts—by proposing an ultra-lightweight structured framework. Built upon Retinex-based residual decomposition, the method uniquely integrates multi-color-space priors within a lightweight architecture, separately optimizing illumination and reflectance components to achieve exposure correction, with an emphasis on enhancement rather than reconstruction. By leveraging multi-prior representations, lightweight neural operations, and multi-representation fusion, the authors construct an efficient network containing only 0.7K–45K parameters. Extensive experiments demonstrate that the proposed approach surpasses existing lightweight state-of-the-art methods across multiple benchmarks and achieves performance comparable to much heavier models, effectively balancing computational efficiency and visual enhancement quality.

This work addresses the challenges of low-light image enhancement, which is prone to noise amplification, artifacts, and color distortion, and often relies solely on RGB data, limiting its ability to model illumination and scene structure effectively. To overcome these limitations, the authors propose M2Retinexformer, a novel framework that integrates depth, luminance priors, and semantic features into the Retinex model for the first time. It introduces a multi-scale cross-attention fusion mechanism and an adaptive gating strategy guided by the reliability of auxiliary cues, dynamically balancing self-attention and cross-modal attention to enable progressive enhancement. Extensive experiments demonstrate that M2Retinexformer significantly outperforms Retinexformer and other state-of-the-art methods across multiple benchmarks—including LOL, SID, SMID, and SDSD—achieving superior performance in brightness restoration, color fidelity, and detail reconstruction.

This work addresses the heteroscedastic noise amplification inherent in Retinex-based low-light image enhancement by proposing an unsupervised, non-learning method. The approach integrates bright channel prior–based illumination estimation, Retinex reflectance decomposition, and edge-preserving denoising, and—within a Retinex framework—introduces for the first time a conditional negative binomial pseudo-count model to characterize the over-dispersed noise induced by division operations. It further provides a boundary-constrained maximum likelihood solution for zero-valued observations, eliminating the need for sensor calibration or deep learning. Experimental results demonstrate state-of-the-art performance among traditional methods on the LOL-v1 dataset, achieving 17.74 dB PSNR and 0.739 SSIM, while attaining real-time processing at 43 FPS on an Apple M2 Pro for 600×400 resolution images.