Deep Retinex Decomposition for Low-Light Enhancement

By Chen Wei, Wenjing Wang, Wenhan Yang, Jiaying Liu. British Machine Vision Conference, 2018

2,652 citations recorded by the source corpus.

Retinex model is an effective tool for low-light image enhancement. It assumes that observed images can be decomposed into the reflectance and illumination. Most existing Retinex-based methods have carefully designed hand-crafted constraints and parameters for this highly ill-posed decomposition, which may be limited by model capacity when applied in various scenes. In this paper, we collect a LOw-Light dataset (LOL) containing low/normal-light image pairs and propose a deep Retinex-Net learned on this dataset, including a Decom-Net for decomposition and an Enhance-Net for illumination adjustment. In the training process for Decom-Net, there is no ground truth of decomposed reflectance and illumination. The network is learned with only key constraints including the consistent reflectance shared by paired low/normal-light images, and the smoothness of illumination. Based on the decomposition, subsequent lightness enhancement is conducted on illumination by an enhancement network called Enhance-Net, and for joint denoising there is a denoising operation on reflectance. The Retinex-Net is end-to-end trainable, so that the learned decomposition is by nature good for lightness adjustment. Extensive experiments demonstrate that our method not only achieves visually pleasing quality for low-light enhancement but also provides a good representation of image decomposition.

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/c7798bbbcd2e85593a2b2d37b92dabac63420cd8

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

Learning Enriched Features for Real Image Restoration and Enhancement (2020, methodology): “ we demonstrate the eectiveness of our algorithm by evaluating it for the image enhancement task. We report PSNR/SSIM values of our method and several other techniques in Table5and Table6for the LoL [107] and MITAdobe FiveK [12] datasets, respectively. It can be seen that our MIRNet achieves signica”
OmniGen: Unified Image Generation (2024, positive): “For low-level vision tasks (low-light image enhancement [67], derain-ing [73], deblurring [44], inpainting [52], outpainting [52] and colorization [57]), where the annotation itself is an image, we only add text instructions, which are randomly sampled from instructions generated by GPT-4o.”
URetinex-Net: Retinex-based Deep Unfolding Network for Low-light Image Enhancement (2022, methodology): “Among these learning-based methods, Retinex-based ones [34,39,41] use deep networks to estimate reﬂectance and illumination, and brighten up illumination.”
Low-Light Image Enhancement with Normalizing Flow (2021, methodology): “The results evaluated on LOL dataset (Wei et al. 2018) are reported in Table 5.”
Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement (2020, methodology): “RetinexNet [27] produces over-exposed artifacts.”
Low-Light Image Enhancement With Semi-Decoupled Decomposition (2020, methodology): “5, we show the results based our decomposition model and other models [4], [6], [21], which also obtain the illumination layer and the reﬂectance layer for low-light enhancement.”
EnlightenGAN: Deep Light Enhancement Without Paired Supervision (2019, methodology): “The LOL dataset [5] is so far the only dataset of low/normal-light image pairs taken from real scenes by changing exposure time and ISO.”
Attention Guided Low-Light Image Enhancement with a Large Scale Low-Light Simulation Dataset (2019, methodology): “map [28] [77]”
Advancing Image Understanding in Poor Visibility Environments: A Collective Benchmark Study (2019, background): “Multi-Scale Retinex (MSR) [105](14); (f) Joint Enhancement and Denoising (JED) [21]16; (g) RetinexNet [59]17.”
MAXIM: Multi-Axis MLP for Image Processing (2022, methodology): “The top row shows an example from the MIT-Adobe FiveK dataset [7], while the bottom row exemplifies a comparison from LOL [96].”
Ultra-High-Definition Low-Light Image Enhancement: A Benchmark and Transformer-Based Method (2022, methodology): “Most existing datasets, e.g., LOL (Wei et al. 2018) and SID (Chen et al. 2018), consist of lower resolution images (1K or less).”
LR3M: Robust Low-Light Enhancement via Low-Rank Regularized Retinex Model (2020, background): “In [26], Wang et al. proposes a deep Retinex-Net including a Decom-Net for decomposition and an Enhance-Net for illumination adjustment.”
Sparse Gradient Regularized Deep Retinex Network for Robust Low-Light Image Enhancement (2021, methodology): “This article is an extension of our previous conference paper [20].”
Benchmarking Low-Light Image Enhancement and Beyond (2021, methodology): “Because previous works, e.g. LOL (Wei et al. 2018), have made great efforts in building datasets similar to VE-LOL-L, in the next part, we only focus on illustrating VE-LOL-H in detail, which is largely beyond considerations of previous works.”
From Fidelity to Perceptual Quality: A Semi-Supervised Approach for Low-Light Image Enhancement (2020, methodology): “We partition LOL dataset [30] for training and testing.”
Toward Fast, Flexible, and Robust Low-Light Image Enhancement (2022, neutral): “…four recently-proposed model-based methods (including LECARM [17], SDD [10], STAR [26]), four advanced supervised learning meth-ods (including RetinexNet [3], KinD [34], FIDE [27], DRBN [30]), and four unsupervised learning methods (including EnGAN [11], SSIENet [33], ZeroDCE [7], and RUAS [14]).”
Joint Rain Detection and Removal from a Single Image with Contextualized Deep Networks (2020, background): “There are also some recent works on bad weather restoration or image enhancement, such as dehazing [36], [37], raindrop and dirt removal [38], [39], light enhancement [40], [41] and moderate rain removal [1], [22], [42], [43].”
Learning Enriched Features for Fast Image Restoration and Enhancement (2022, methodology): “Method BIMEF CRM Dong LIME MF RRM SRIE Retinex-Net MSR NPE GLAD KinD KinD++ MIRNet-v2 [126] [127] [128] [129] [130] [131] [130] [85] [81] [132] [133] [4] [134] (Ours)”