PaperlayerView exact revision on arXiv

Source-linked AI summary

Kindling the Darkness: A Practical Low-light Image Enhancer

Yonghua Zhang, Jiawan Zhang, Xiaojie Guo

arXiv:1905.04161v1cs.CV

TL;DR

Low-light enhancement must improve visibility without amplifying hidden noise and color distortion, yet lacks a single ground-truth light level. KinD separates illumination and reflectance, trains from paired exposures, restores degraded reflectance, and flexibly adjusts illumination. On LOL, KinD significantly outperforms competing methods on PSNR, SSIM, and NIQE, while visual comparisons show effective defect removal and light adjustment.

  • Problem

    Low-light images combine poor visibility with hidden noise and color distortion, while no single ground-truth light condition fits diverse user requirements.

  • Method

    KinD uses Retinex-inspired reflectance–illumination decomposition, paired-exposure training, illumination-guided reflectance restoration, and a user-controlled illumination adjustment function.

  • Results

    KinD significantly outperforms other methods on LOL's PSNR, SSIM, and NIQE evaluations, while visual comparisons show adjusted light and removed degradations.

  • Takeaways & Limitations

    KinD provides a practical enhancer that removes visual defects while allowing flexible adjustment of light levels.

  • Takeaways & Limitations

    The method operates without well-defined ground-truth light conditions or abundant ground-truth reflectance and illumination maps for real images.

Abstract

from arXiv · show

Images captured under low-light conditions often suffer from (partially) poor visibility. Besides unsatisfactory lightings, multiple types of degradations, such as noise and color distortion due to the limited quality of cameras, hide in the dark. In other words, solely turning up the brightness of dark regions will inevitably amplify hidden artifacts. This work builds a simple yet effective network for \textbf{Kin}dling the \textbf{D}arkness (denoted as KinD), which, inspired by Retinex theory, decomposes images into two components. One component (illumination) is responsible for light adjustment, while the other (reflectance) for degradation removal. In such a way, the original space is decoupled into two smaller subspaces, expecting to be better regularized/learned. It is worth to note that our network is trained with paired images shot under different exposure conditions, instead of using any ground-truth reflectance and illumination information. Extensive experiments are conducted to demonstrate the efficacy of our design and its superiority over state-of-the-art alternatives. Our KinD is robust against severe visual defects, and user-friendly to arbitrarily adjust light levels. In addition, our model spends less than 50ms to process an image in VGA resolution on a 2080Ti GPU. All the above merits make our KinD attractive for practical use.

1 INTRODUCTION

Low-light enhancement must address poor visibility alongside noise, color distortion, and saturation, while accommodating the absence of a single ground-truth light level. KinD addresses these challenges with Retinex-inspired decomposition, paired-image training, degradation removal, and flexible illumination adjustment.

  • High ISO, long exposure, and flash each introduce practical drawbacks, including amplified noise, blur, or unbalanced lighting.
  • Simply amplifying dark regions reveals hidden noise and color distortion, while bright regions can become saturated or overexposed.
  • Low-light enhancement lacks a single ground-truth light condition because preferred illumination levels vary across users and applications.
  • The central challenges are estimating and flexibly adjusting illumination, removing hidden degradations, and training without well-defined target light conditions.
  • KinD decomposes images into reflectance and illumination, trains on paired exposures without ground-truth layers, and includes modules for flexible adjustment and defect removal.
  • Extensive experiments are used to demonstrate KinD's efficacy and superiority over state-of-the-art alternatives.

2 METHODOLOGY

KinD uses separate branches for reflectance and illumination, organized functionally into decomposition, reflectance restoration, and illumination adjustment modules.

  • KinD uses two branches to handle reflectance and illumination components, respectively.
  • Functionally, the network comprises layer decomposition, reflectance restoration, and illumination adjustment modules.

2.1 Consideration & Motivation

KinD uses Retinex decomposition to separate illumination from reflectance and guide restoration, while addressing underdetermination and flexible user-controlled lighting.

  • Retinex models an image as I = R ◦ L, separating reflectance from illumination into two smaller mapping subspaces.
  • Because real images lack well-defined ground-truth light, reflectance, and illumination maps, decomposition requires additional priors and regularizers.
  • Paired exposures support shared-reflectance regularization and mutually consistent, piece-wise smooth illumination constraints.
  • The degraded low-light image is modeled as I = R ◦ L + E, with illumination-dependent degradation represented in the reflectance component.
  • Illumination guides reflectance restoration because degradation varies with illumination and cannot be uniformly removed across the image.
  • A practical enhancer needs arbitrary illumination manipulation because preferred light strengths differ across people and applications.

2.2 KinD Network

KinD decomposes low-light images into reflectance and illumination, then separately restores reflectance and adjusts illumination. It is trained using paired exposures and supports flexible light manipulation through three functional modules.

  • 2.2 KinD Network: KinD comprises layer decomposition, reflectance restoration, and illumination adjustment modules for handling reflectance and illumination components.The network has separate branches for reflectance and illumination.
  • 2.2 KinD Network: Paired images with different exposures regularize reflectance similarity, illumination smoothness, mutual consistency, and reconstruction without ground-truth layers.The decomposition uses shared-scene reflectance and structured illumination assumptions.
  • 2.2 KinD Network: The decomposition network uses a U-Net reflectance branch and an illumination branch built from convolutional layers on concatenated reflectance features.The illumination branch can potentially exclude textures from illumination.
  • 2.2 KinD Network: The reflectance restoration net uses clearer reflectance from brighter images as an informal reference and incorporates illumination information because degradation depends on illumination distribution.Figure 5 compares polluted reflectance, BM3D results, and the restoration network under different degradation levels.
  • 2.2 KinD Network: The illumination adjustment net estimates α from paired illumination maps and learns to convert a source illumination toward a target condition.The ratio α is computed from paired target and source illuminations and supplied as a feature-map input.
  • 2.2 KinD Network: Compared with gamma correction, KinD adjusts bright and dark regions differently and lets users control light strength directly through α.Setting α to 2 means turning the light 2x up.

3 EXPERIMENTAL VALIDATION

KinD is evaluated on multiple low-light datasets against state-of-the-art methods using reference and nonreference metrics, quantitative comparisons, and visual examples. It achieves strong results overall, while LOE evaluation depends on the choice of reference.

  • The model is trained on 450 of LOL’s 500 paired low/normal-light images, without synthetic images.Training uses SGD, with 48x48 patches for decomposition and 384x384 patches for reflectance restoration and illumination adjustment.
  • KinD significantly outperforms competing methods on LOL in PSNR and SSIM and ranks first by a large margin in NIQE.For LOEref, KinD ranks third, slightly behind CRM at 977.3 versus 926.1.
  • KinD shows a clear NIQE advantage on the LIME and NPE datasets.
  • 3.34 vs. 3.27: KinD trails CRM by a small margin on the MEF dataset’s NIQE evaluation.
  • Visual comparisons show that KinD properly adjusts light and clearly removes defects that remain in competing methods.The reported residual defects include unsatisfactory illumination adjustment, noise, and color distortion.

4 CONCLUSION

KinD decomposes low-light images into reflectance and illumination, trains with paired images under different exposures, and separately restores degradation and adjusts illumination. Experiments report advantages over state-of-the-art alternatives, with flexible light control and VGA processing under 50ms on an Nvidia 2080Ti GPU.

  • KinD decomposes images into reflectance and illumination, decoupling the original space into two smaller subspaces.
  • The network is alternatively trained with paired images captured under different light or exposure conditions, without ground-truth reflectance and illumination information.
  • KinD includes restoration and learned illumination-adjustment modules to remove hidden degradations and flexibly adjust light levels.
  • Less than 50ms: KinD processes a VGA-resolution image on an Nvidia 2080Ti GPU.
Loading 1905.04161v1…