Back

Deep Color Consistent Network for Low-Light Image Enhancement

Zhao ZhangHuan ZhengRichang HongMingliang XuShuicheng YanMeng Wang

article2022CVPR184 citations

Proposes DCC-Net, a low-light image restoration framework that decouples images into structural gray components and color histograms to eliminate color distortion through pyramid feature embedding.

Listen

Low-light image enhancement is essential for improving visual quality and supporting downstream computer vision applications, such as object detection, face recognition, and semantic segmentation. While deep learning methods have significantly advanced image brightening, existing approaches predominantly focus on adjusting illumination. Consequently, they often overlook color fidelity, producing unnatural images with severe color distortion and noticeable gaps compared to ground-truth normal-light scenes.

The article evaluates a new deep neural network architecture called the Deep Color Consistent Network (DCC-Net). The primary objective is to demonstrate that directly incorporating color preservation alongside illumination enhancement can eliminate color discrepancies and generate more natural, visually coherent results.

To achieve this, the authors introduce a collaborative strategy that separates the enhancement task into distinct components. One sub-network reconstructs a grayscale image to capture scene structure and texture, while a second sub-network learns global color distributions via color histograms. A third sub-network integrates these two elements using a multi-level Pyramid Color Embedding module, which applies a dual affinity matrix to dynamically match color features with spatial content. The authors evaluated DCC-Net against six leading deep enhancement models across six standard public datasets, using both paired reference data and real-world unpaired imagery under diverse lighting conditions.

The key findings demonstrate significant performance gains over current state-of-the-art techniques. DCC-Net achieved the lowest color distortion on standard benchmark testing, reducing color error substantially compared to competing methods, some of which exhibited up to nine times higher error rates. The model achieved superior image fidelity and structural clarity metrics, recording the lowest reconstruction error and highest overall signal quality. Visual assessments confirmed that DCC-Net effectively avoids the common over-enhancement, washed-out whites, and unnatural color casts typical of competing models. Furthermore, DCC-Net maintained practical computational efficiency, processing images in approximately 0.026 seconds per frame, making it faster than several supervised baselines.

These results demonstrate that separating structural recovery from color estimation successfully overcomes the information mismatch that degrades traditional enhancement methods. For decision-makers and engineering teams, this approach offers a dependable foundation for vision systems operating in low-light conditions, mitigating the risk of downstream processing failures caused by distorted visual inputs.

For future implementation and research, the authors recommend exploring further architectural refinements to boost naturalness and developing standardized, quantitative assessment metrics specifically tailored to content naturalness and color discrepancy. A current limitation is that testing relied on established academic benchmark datasets, and quantitatively measuring perceptual naturalness remains an evolving research challenge. Nevertheless, confidence in the findings is high, strongly supported by consistent quantitative advantages across multiple datasets and distinct ablation studies confirming the impact of each network module.

color differencecolor histogramreal data setsDeep Color Consistent Networklow-light image enhancementcontrast enhancement
Zhang et al (2022).pdf
  • Paper: HVI: A New Color Space for Low-light Image Enhancement, Qingsen Yan et al. (2025). HVI carries forward the challenge of color distortion in low-light enhancement by proposing a color-space representation designed to separate chromatic information from brightness.
Cover for Deep Color Consistent Network for Low-Light Image Enhancement

Abstract

Low-light image enhancement (LLIE) explores how to refine the illumination and obtain natural normal-light images. Current LLIE methods mainly focus on improving the illumination, but do not consider the color consistency by reasonably incorporating color information into the LLIE process. As a result, color difference usually exists between the enhanced image and ground-truth. To address this issue, we propose a new deep color consistent network termed DCC-Net to retain the color consistency for LLIE. A new “divide and conquer” collaborative strategy is presented, which can jointly preserve color information and enhance the illumination. Specifically, the decoupling strategy of our DCC-Net decouples each color image into two main components, i.e., gray image plus color histogram. Gray image is used to generate reasonable structures and textures, and the color histogram is beneficial for preserving the color consistency. That is, they both are utilized to complete the LLIE task collaboratively. To match the color and content features, and reduce the color consistency gap between enhanced image and ground-truth, we also design a new pyramid color embedding (PCE) module, which can better embed color information into the LLIE process. Extensive experiments on six real datasets show that the enhanced images of our DCC-Net are more natural and colorful, and perform favorably against the state-of-the-art methods.

Topics

color differencecolor histogramreal data setscolor componentdigital imagereference imagecolor features

Keywords

Deep Color Consistent Networklow-light image enhancementcontrast enhancementconsistent color featuresillumination mapimage reconstruction lossdeep autoencodercolor histogram

Table of Contents

  • 1. Introduction
  • 2. Related work
  • 2.1. Traditional LLIE Methods
  • 2.2. Deep Learning-based LLIE Methods
  • 3. Proposed Method
  • 3.1. Network Structure
  • 3.2. Pyramid Color Embedding (PCE)
  • 3.3. Objective Function
  • 4. Experiments
  • 4.1. Experimental Settings
  • 4.2. Quantitative Enhancement Results
  • 4.3. Visual Image Analysis and Evaluations
  • 4.4. Ablation Study
  • 5. Conclusion
  • 6. Acknowledgments
  • References

Knowls

  1. Knowl 1 — Decoupling Strategy and Three-Subnet Architecture of DCC-Net

    model/method

    DCC-Net is a deep neural network architecture for low-light image enhancement (LLIE) that addresses color inconsistency by decoupling an input low-light color image Slow∈R3×H×WS_{low} \in \mathbb{R}^{3 \times H \times W} into structural content and global color distribution components through three collaborative sub-networks:

    1. G-Net (Gray Network): An encoder-decoder network (based on U-Net) that predicts a normal-light grayscale image Gpre=GNet(Slow)G_{pre} = \text{GNet}(S_{low}), where Gpre∈R1×H×WG_{pre} \in \mathbb{R}^{1 \times H \times W}. G-Net recovers textures, edge structures, and illumination without color distractions.
    2. C-Net (Color Network): An encoder-decoder network that maps SlowS_{low} directly to a predicted RGB color histogram matrix Cpre=CNet(Slow)C_{pre} = \text{CNet}(S_{low}), where Cpre∈R3×256C_{pre} \in \mathbb{R}^{3 \times 256}. Each row of CpreC_{pre} corresponds to one of the three color channels (R,G,BR, G, B) across 256 pixel intensity bins, learning spatial-invariant global color distributions.
    3. R-Net (Restoration Network): A synthesis network that combines the original input SlowS_{low}, the predicted gray image GpreG_{pre}, and the predicted color histogram CpreC_{pre} to reconstruct the final enhanced normal-light color image Spre=RNet(Slow,Gpre,Cpre)S_{pre} = \text{RNet}(S_{low}, G_{pre}, C_{pre}), where Spre∈R3×H×WS_{pre} \in \mathbb{R}^{3 \times H \times W}. R-Net incorporates the Pyramid Color Embedding (PCE) module to progressively integrate color features into decoder layers.

  2. Knowl 2 — Dual Affinity Matrix for Content-Color Feature Alignment

    equation

    To overcome the absence of spatial information in global color histograms and prevent feature mismatch between color attributes and spatial contents, a Dual Affinity Matrix (DAM) computes pixel-wise spatial affinities between content feature maps F∈RN×H×WF \in \mathbb{R}^{N \times H \times W} (from the encoder of R-Net) and color feature maps C∈RN×H×WC \in \mathbb{R}^{N \times H \times W} (derived from the color histogram):

    M(x, y) &= -\|F(x, y) - C(x, y)\|_1 \\ P(x, y) &= F(x, y) \cdot C(x, y) \\ A &= 2 \times \text{sigmoid}(M) \odot \tanh(P) \end{aligned}$$ where: - $(x, y)$ denotes the spatial pixel coordinate with $x \in \{1, \dots, H\}$ and $y \in \{1, \dots, W\}$. - $F(x, y), C(x, y) \in \mathbb{R}^N$ are the feature vectors at position $(x, y)$ across channel dimension $N$. - $M \in \mathbb{R}^{H \times W}$ is the negative Manhattan distance matrix. Because $M(x, y) \le 0$ everywhere, $\text{sigmoid}(M(x, y)) \in [0, 0.5]$, ensuring $2 \times \text{sigmoid}(M) \in [0, 1]$. - $P \in \mathbb{R}^{H \times W}$ is the inner product matrix capturing directional feature correlation. - $\odot$ denotes element-wise multiplication. - $A \in [0, 1]^{H \times W}$ is the dual affinity matrix that dynamically weights color feature embedding.

  3. Knowl 3 — Pyramid Color Embedding Module

    model/method

    The Pyramid Color Embedding (PCE) module dynamically injects color distribution features into the decoder stages of the restoration network (R-Net) across six hierarchical resolution levels.

    A single Color Embedding (CE) sub-module takes an encoder content feature map F∈RN×H×WF \in \mathbb{R}^{N \times H \times W} and a color feature map C∈RN×H×WC \in \mathbb{R}^{N \times H \times W}, computes the dual affinity matrix A∈[0,1]H×WA \in [0, 1]^{H \times W}, and calculates the color-embedded feature map EE:

    E=A⊙C+FE = A \odot C + F

    where ⊙\odot denotes element-wise multiplication broadcast across channels.

    Across the 6-layer pyramid structure:

    Ei,Ci+1=CE(Fi,Ci),i=1,2,…,6E_i, C_{i+1} = \text{CE}(F_i, C_i), \quad i = 1, 2, \dots, 6

    where FiF_i is copied from the ii-th encoder layer of R-Net, C1C_1 is initialized from the color feature output of C-Net, EiE_i is supplied to the corresponding decoder layer of R-Net, and Ci+1C_{i+1} is produced by upsampling the spatial resolution of CiC_i before feeding into the (i+1)(i+1)-th CE sub-module.

  4. Knowl 4 — Multi-Term Joint Loss Function for DCC-Net

    equation

    DCC-Net is trained end-to-end using a composite objective function ltotall_{total} that constrains grayscale reconstruction, color histogram fidelity, pixel-level color restoration, structural similarity, and surface smoothness:

    ltotal=λglg+λclc+λrlr+λssimlssim+λtvltvl_{total} = \lambda_g l_g + \lambda_c l_c + \lambda_r l_r + \lambda_{ssim} l_{ssim} + \lambda_{tv} l_{tv}

    where the loss terms and their trade-off hyperparameters are:

    1. Grayscale Reconstruction Loss (λg=1\lambda_g = 1): lg=1H×W∥Gpre−Ghigh∥1l_g = \frac{1}{H \times W} \|G_{pre} - G_{high}\|_1 where GhighG_{high} is the grayscale representation of the normal-light ground truth.

    2. Color Histogram Reconstruction Loss (λc=2\lambda_c = 2): lc=1N×256∥Cpre−Chigh∥1l_c = \frac{1}{N \times 256} \|C_{pre} - C_{high}\|_1 where ChighC_{high} is the ground-truth RGB color histogram matrix (N=3N = 3).

    3. Pixel-level Color Reconstruction Loss (λr=2\lambda_r = 2): lr=1N×H×W∥Spre−Shigh∥1l_r = \frac{1}{N \times H \times W} \|S_{pre} - S_{high}\|_1 where ShighS_{high} is the ground-truth normal-light color image.

    4. Structural Similarity Loss (λssim=2\lambda_{ssim} = 2): lssim=1−SSIM(Spre,Shigh)l_{ssim} = 1 - \text{SSIM}(S_{pre}, S_{high}) with SSIM(x,y)=2μxμy+c1μx2+μy2+c1⋅2σxy+c2σx2+σy2+c2\text{SSIM}(x, y) = \frac{2\mu_x\mu_y + c_1}{\mu_x^2 + \mu_y^2 + c_1} \cdot \frac{2\sigma_{xy} + c_2}{\sigma_x^2 + \sigma_y^2 + c_2}, where μx,μy\mu_x, \mu_y are means, σx2,σy2\sigma_x^2, \sigma_y^2 are variances, σxy\sigma_{xy} is covariance, and c1,c2c_1, c_2 are stabilization constants.

    5. Total Variation Regularization Loss (λtv=0.1\lambda_{tv} = 0.1): ltvl_{tv} enforces spatial smoothness on SpreS_{pre} to prevent overfitting.

  5. Knowl 5 — Experimental Setup and Implementation Details for DCC-Net

    experimental setup

    DCC-Net was implemented in PyTorch and evaluated using the following specifications:

    • Training Datasets: 1,000 synthetic paired low/normal-light image pairs from the LOL synthetic dataset plus 485 real low/normal-light image pairs from the LOL real dataset.
    • Testing Datasets:
      • Paired real benchmark: LOL test dataset (15 paired images).
      • Unpaired real benchmarks: DICM (64 images), LIME (10 images), MEF (17 images), NPE (85 images), and VV (24 images).
    • Evaluation Metrics:
      • Full-reference: Peak Signal-to-Noise Ratio (PSNR in dB), Structural Similarity Index (SSIM), Mean Absolute Error (MAE in percentage), and Color-Sensitive Error ratio (CSE ratio relative to DCC-Net).
      • Non-reference: Naturalness Image Quality Evaluator (NIQE).
      • Efficiency: Inference time per image in seconds.
    • Optimization and Training: Adam optimizer with batch size 6 on two NVIDIA GeForce RTX 2080 Ti GPUs. All training and testing images are resized to 512×512512 \times 512 pixels. Training runs for 400 epochs with a learning rate of 0.00010.0001 for the first 200 epochs and 0.000010.00001 for the final 200 epochs.

  6. Knowl 6 — Quantitative Evaluation on the Paired LOL Dataset

    data/table

    DCC-Net was quantitatively compared against six deep LLIE methods on the 15 paired test images of the LOL dataset. The compared methods include RetinexNet, KinD, Zero-DCE, EnlightenGAN, Zero-DCE++, and KinD++. Metrics evaluated are PSNR (dB, higher is better), SSIM (higher is better), MAE (%, lower is better), CSE ratio (ratio of method error to DCC-Net error, lower is better), and inference time (s).

    Method PSNR SSIM MAE(%) CSE (ratio) Time(s)
    RetinexNet 16.82 0.43 14.93 2.51 0.0390
    KinD 20.42 0.82 9.82 2.30 0.0650
    Zero-DCE 16.02 0.51 15.98 5.98 0.0026
    EnlightenGAN 18.32 0.64 13.71 1.66 0.0150
    Zero-DCE++ 16.11 0.53 15.89 9.59 0.0012
    KinD++ 20.92 0.80 8.83 1.15 0.0320
    DCC-Net 22.72 0.81 8.72 1.00 0.0260

    DCC-Net achieves the best PSNR (22.72 dB22.72\text{ dB}), lowest MAE (8.72%8.72\%) and lowest color error (CSE ratio 1.001.00), while obtaining a competitive SSIM of 0.810.81 and an inference time of 0.0260 s0.0260\text{ s} per 512×512512 \times 512 image.

  7. Knowl 7 — Non-Reference Naturalness Evaluation on Unpaired Low-Light Datasets

    data/table

    To evaluate generalization on real low-light images without ground-truth counterparts, DCC-Net was benchmarked against six deep LLIE methods across five unpaired datasets (DICM, LIME, MEF, NPE, VV) using the Naturalness Image Quality Evaluator (NIQE) metric, where lower values indicate better perceptual naturalness and image quality.

    Method DICM LIME MEF NPE VV
    RetinexNet 4.33 5.75 4.93 4.95 4.32
    KinD 3.95 4.42 4.45 3.92 3.72
    Zero-DCE 4.58 5.82 4.93 4.53 4.81
    EnlightenGAN 4.06 4.59 4.70 3.99 4.04
    Zero-DCE++ 4.89 5.66 5.10 4.74 5.10
    KinD++ 3.89 4.90 4.55 3.91 3.82
    DCC-Net 3.70 4.42 4.59 3.70 3.28

    DCC-Net achieves the best (lowest) NIQE scores on DICM (3.703.70), NPE (3.703.70), and VV (3.283.28), matches the best performance of KinD on LIME (4.424.42), and achieves competitive performance on MEF (4.594.59).

  8. Knowl 8 — Ablation Analysis of DCC-Net Sub-Networks and PCE Module

    data/table

    An ablation study on the LOL dataset isolates the impact of each core structural component in DCC-Net:

    • W/o G-Net: DCC-Net without the grayscale structure reconstruction sub-network G-Net.
    • W/o C-Net: DCC-Net without the color histogram learning sub-network C-Net.
    • W/o PCE: DCC-Net replacing the Pyramid Color Embedding module with simple feature concatenation.
    • DCC-Net: The full proposed network.
    Model W/o G-Net W/o C-Net W/o PCE DCC-Net
    PSNR (dB) 21.51 21.01 21.14 22.72
    SSIM 0.79 0.79 0.79 0.81
    MAE (%) 10.27 10.13 10.43 8.72

    Removing any individual component leads to noticeable performance drops: omitting C-Net causes the steepest PSNR reduction (down to 21.01 dB21.01\text{ dB}), while omitting PCE reduces PSNR to 21.14 dB21.14\text{ dB} and increases MAE to 10.43%10.43\%, confirming that both dedicated color distribution modeling and affinity-guided pyramid embedding are required for color consistency.

Coverage note — No substantial contributed material was omitted. All primary architectural contributions (G-Net, C-Net, R-Net, DAM, PCE), the multi-task loss objective, experimental protocols, quantitative benchmark comparisons, and ablation studies are fully represented.

References

  1. 1.Turgay Celik and Tardi Tjahjadi. Contextual and variational contrast enhancement. IEEE Transactions on Image Processing, 20(12):3431–3441, 2011.
  2. 2.Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3291–3300, 2018.
  3. 3.Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690–4699, 2019.
  4. 4.Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3146–3154, 2019.
  5. 5.Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014.
  6. 6.Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1780–1789, 2020.
  7. 7.Xiaojie Guo. Lime: a method for low-light image enhancement. In Proceedings of the 24th ACM international conference on Multimedia, pages 87–91, 2016.
  8. 8.Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE Transactions on image processing, 26(2):982–993, 2016.
  9. 9.Ju Han and Kai-Kuang Ma. Fuzzy color histogram and its use in color image retrieval. IEEE Transactions on image Processing, 11(8):944–952, 2002.
  10. 10.Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.
  11. 11.Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhancement without paired supervision. IEEE Transactions on Image Processing, 30:2340–2349, 2021.
  12. 12.Daniel J Jobson, Zia-ur Rahman, and Glenn A Woodell. A multiscale retinex for bridging the gap between color images and the human observation of scenes. IEEE Transactions on Image processing, 6(7):965–976, 1997.
  13. 13.Daniel J Jobson, Zia-ur Rahman, and Glenn A Woodell. Properties and performance of a center/surround retinex. IEEE transactions on Image Processing, 6(3):451–462, 1997.
  14. 14.Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
  15. 15.Edwin H Land. The retinex theory of color vision. Scientific american, 237(6):108–129, 1977.
  16. 16.Chulwoo Lee, Chul Lee, and Chang-Su Kim. Contrast enhancement based on layered difference representation. In 2012 19th IEEE International Conference on Image Processing, pages 965–968. IEEE, 2012.
  17. 17.Chulwoo Lee, Chul Lee, and Chang-Su Kim. Contrast enhancement based on layered difference representation of 2d histograms. IEEE transactions on image processing, 22(12):5372–5384, 2013.
  18. 18.Chongyi Li, Chunle Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. arXiv preprint arXiv:2103.00860, 2021.
  19. 19.Chongyi Li, Jichang Guo, Fatih Porikli, and Yanwei Pang. Lightennet: A convolutional neural network for weakly illuminated image enhancement. Pattern recognition letters, 104:15–22, 2018.
  20. 20.Mading Li, Jiaying Liu, Wenhan Yang, Xiaoyan Sun, and Zongming Guo. Structure-revealing low-light image enhancement via robust retinex model. IEEE Transactions on Image Processing, 27(6):2828–2841, 2018.
  21. 21.Kin Gwn Lore, Adedotun Akintayo, and Soumik Sarkar. Llnet: A deep autoencoder approach to natural low-light image enhancement. Pattern Recognition, 61:650–662, 2017.
  22. 22.Kede Ma, Kai Zeng, and Zhou Wang. Perceptual quality assessment for multi-exposure image fusion. IEEE Transactions on Image Processing, 24(11):3345–3356, 2015.
  23. 23.Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems, 32:8026–8037, 2019.
  24. 24.Stephen M Pizer. Contrast-limited adaptive histogram equalization: Speed and effectiveness stephen m. pizer, r. eugene johnston, james p. ericksen, bonnie c. yankaskas, keith e. muller medical image display research group. In Proceedings of the First Conference on Visualization in Biomedical Computing, Atlanta, Georgia, volume 337, 1990.
  25. 25.Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28:91–99, 2015.
  26. 26.Wenqi Ren, Sifei Liu, Lin Ma, Qianqian Xu, Xiangyu Xu, Xiaochun Cao, Junping Du, and Ming-Hsuan Yang. Low-light image enhancement via a deep hybrid network. IEEE Transactions on Image Processing, 28(9):4364–4375, 2019.
  27. 27.Xutong Ren, Wenhan Yang, Wen-Huang Cheng, and Jiaying Liu. Lr3m: Robust low-light enhancement via low-rank regularized retinex model. IEEE Transactions on Image Processing, 29:5862–5876, 2020.
  28. 28.Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015.
  29. 29.Shuhang Wang, Jin Zheng, Hai-Miao Hu, and Bo Li. Naturalness preserved enhancement algorithm for non-uniform illumination images. IEEE Transactions on Image Processing, 22(9):3538–3548, 2013.
  30. 30.Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retinex decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018.
  31. 31.Yanyan Wei, Zhao Zhang, Yang Wang, Mingliang Xu, Yi Yang, Shuicheng Yan, and Meng Wang. Deraincyclegan: Rain attentive cyclegan for single image deraining and rainmaking. IEEE Transactions on Image Processing, 30:4788–4801, 2021.
  32. 32.Wenhan Yang, Shiqi Wang, Yuming Fang, Yue Wang, and Jiaying Liu. From fidelity to perceptual quality: A semi-supervised approach for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3063–3072, 2020.
  33. 33.Lin Zhang, Lijun Zhang, Xiao Liu, Ying Shen, Shaoming Zhang, and Shengjie Zhao. Zero-shot restoration of back-lit images using deep internal learning. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1623–1631, 2019.
  34. 34.Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision, 129(4):1013–1037, 2021.
  35. 35.Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the 27th ACM international conference on multimedia, pages 1632–1640, 2019.
  36. 36.Zhao Zhang, Yanyan Wei, Haijun Zhang, Yi Yang, Shuicheng Yan, and Meng Wang. Data-driven single image deraining: A comprehensive review and new perspectives. TechRxiv Preprint, pages 1–20, 2021.
  37. 37.Suiyi Zhao, Zhao Zhang, Richang Hong, Mingliang Xu, Haijun Zhang, Meng Wang, and Shuicheng Yan. Unsupervised color retention network and new quantization metric for blind motion deblurring. TechRxiv Preprint, 2021.
  38. 38.Anqi Zhu, Lin Zhang, Ying Shen, Yong Ma, Shengjie Zhao, and Yicong Zhou. Zero-shot restoration of underexposed images via robust retinex decomposition. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6. IEEE, 2020.

Citation

MLA
Zhang, Z., et al. “Deep Color Consistent Network for Low-Light Image Enhancement”. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 1889–98, https://doi.org/10.1109/CVPR52688.2022.00194.
APA
Zhang, Z., Zheng, H., Hong, R., Xu, M., Yan, S., & Wang, M. (2022). Deep Color Consistent Network for Low-Light Image Enhancement. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1889–1898. https://doi.org/10.1109/CVPR52688.2022.00194
Chicago
Zhang, Z., H. Zheng, R. Hong, M. Xu, S. Yan, and M. Wang. 2022. “Deep Color Consistent Network for Low-Light Image Enhancement”. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1889–98. https://doi.org/10.1109/CVPR52688.2022.00194.
Harvard
Zhang, Z. et al. (2022) “Deep Color Consistent Network for Low-Light Image Enhancement”, 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp. 1889–1898. Available at: https://doi.org/10.1109/CVPR52688.2022.00194.
Vancouver
1. Zhang Z, Zheng H, Hong R, Xu M, Yan S, Wang M (2022) Deep Color Consistent Network for Low-Light Image Enhancement. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp 1889–1898

BibTeX

@inproceedings{Zhang_2022, title={Deep Color Consistent Network for Low-Light Image Enhancement}, url={http://dx.doi.org/10.1109/CVPR52688.2022.00194}, DOI={10.1109/cvpr52688.2022.00194}, booktitle={2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, publisher={IEEE}, author={Zhang, Zhao and Zheng, Huan and Hong, Richang and Xu, Mingliang and Yan, Shuicheng and Wang, Meng}, year={2022}, month=June, pages={1889–1898} }
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: IEEE