Frontiers in Computing and Intelligent Systems 
ISSN: 2832-6024 | Vol. 1, No. 3, 2022 

 

25 

Low-light image enhancement with contrast 
regularization 

Jinfeng Li* 

College of Computer Science & Technology, Qingdao University, Qingdao 266071, Shandong, China. 

* Corresponding author: Email: yasuierba@163.com. 

 

Abstract: Because the processing of existing low-light images undergoes multiple sampling processing, there is serious 

information degradation, and only clear images are used as positive samples to guide network training, low-light image 

enhancement processing is still a challenging and unsettled problem. Therefore, a multi-scale contrast learning low-light image 

enhancement network is proposed. First, the image generates rich features through the input module, and then the features are 

imported into a multi-scale enhancement network with dense residual blocks, using positive and negative samples to guide the 

network training, and finally using the refinement module to enrich the image details. Experimental results on the dataset show 

that this method can reduce noise and artifacts in low-light images, and can improve contrast and brightness, demonstrating its 

advantages. 

Keywords: Deep Learning; Low-light enhancement; Residual network; Contrast regularization. 

 

1. Introduction 

As an important medium for perceiving the world, the 

quality of images often determines the performance of 

algorithms, and the images taken under low-light conditions 

contain a lot of noise, which affects the information 

acquisition in the application process of high-level visual 

tasks, such as semantic segmentation and target recognition. 

There are many solutions for low-light image enhancement, 

which are divided into traditional methods and deep learning-

based methods according to different algorithmic concepts. 

1.1. Low-light enhancement based on 

traditional methods 

Traditional approaches to low-light image enhancement are 

basically divided into two categories: methods based on 

distribution mapping [1] and methods based on model 

optimization. The distribution mapping method uses curve 

transformation, histogram equalization and other means to 

improve the pixel distribution of the image to improve the 

brightness and clarity of the image, but does not consider 

adjacent pixels, resulting in color distortion and abnormal 

details. The core of the model-based optimization method is 

to take the data derived from the physical imaging law as the 

basic component and further use the existing optimization 

technology to solve. For example, Retinex theory [2,3] based 

on color sensitivity consistency, set the original image S is the 

product of light image L and reflected image R, and then 

decompose R to eliminate the effect of light unevenness and 

improve image quality. However, this method does not 

consider the saturation component and the luminance 

component, and there is still a problem of blurred detail in 

low-light images. 

1.2. Low-light enhancement based on deep 

learning  

In order to overcome the drawbacks of the above traditional 

methods, and with the development of deep learning [4-6], the 

relationship between low-light input and enhanced output has 

become a mainstream low-light image enhancement mode by 

heuristically designing network structures. Two-stage object 

detection algorithms represented by RCNN [7], Fast RCNN 

[8], Faster RCNN [9] and single-stage object detection 

algorithms represented by YOLO [10] and SSD [11] have 

emerged, but these methods focus on individual performance 

enhancements. Detail recovery and color correction for low-

light images are still a huge challenge. 

2. Proposed Method 

2.1. The main network structures 

The approach we propose is the network architecture 

shown in Figure 1, which is an end-to-end model. In order for 

the network to learn more features of multiple dimensions, a 

3*3 convolutional layer is used to increase the number of 

feature dimensions. In order to enable information between 

the different dimensions to communicate, we borrowed the 

idea of up sampling and down sampling in U-Net [12] and 

then used the FA block in FFA-Net [13] with efficient FA 

blocks as our base blocks. Unlike FFA-Net, in order to 

improve computational efficiency and reduce memory 

storage, we reduced the 57 FA blocks in FFA-Net to 6 FA 

blocks. Specifically, a 4x down sampling operation, such as a 

regular convolution in step 1 and two convolutional layers 

with two steps in 2, causes dense FA blocks to learn feature 

representations in low-resolution spaces, and then uses the 

corresponding 4x up sampling and a regular convolution to 

generate the recovered image. 

In Figure 1, we represent the low-light image, the 

corresponding recovered image generated by the low-light 

enhancement network, and its sharp image as negatives, 

anchor points, and positives, respectively. Contrast 

regularization brings the resulting corresponding recovery 

diagram closer to the sharp image and away from the blurry 

image. And the anchor image can be constrained to a closed 

upper and lower bounds by contrast learning, which will help 

the low-light enhancement network to approach the positive 

image and away from the negative image faster. Finally, a 

refinement processing module, the GRDB module, is used to 



 

26 

output the final low-light enhancement image through a 

convolutional 3×3. 

The overall model of the end-to-end single-image low-light 

enhancement approach can be represented as follows: 

arg min
𝑤

‖𝐽 − 𝜙(𝐼, 𝑤)‖ + 𝛽𝜌(𝜙(𝐼, 𝑤))                  (1) 

where I is the low-light image, J is the corresponding sharp 

image, and 𝜙  (·, w) is the low-light enhancement network 

with the parameter w. ‖𝐽 − 𝜙(𝐼, 𝑤)‖ is a data fidelity term 

that typically uses losses based on the L1/L2 norm. 𝜌 (·) is a 

regularization term that produces naturally smooth defogging 

images, where TV norm [14, 15] and DCP prior [14, 15] are 

widely used for regularization terms. β is a penalty parameter 

used to balance the data fidelity and regularization terms. 

 

Figure 1. The network structure diagram of the algorithm in this paper

2.2. Loss function 

In this paper, SSIM loss and Smooth L1 loss are used to 

effectively restore low-light images. 

As SSIM refers to the overall visual effect, it measures the 

difference between the original and enhanced images for the 

purpose of improving their quality. 

𝐿𝑆𝑆𝐼𝑀 = 1 − ∑
2𝜇𝑥𝜇𝑦+𝐶1

𝜇𝑥2+𝜇𝑦2+𝐶1
⋅

2𝜎𝑥𝑦+𝐶2

𝜎𝑥2+𝜎𝑦2+𝐶2

𝑀
𝑚=1   (2) 

Smooth L1 loss can limit the gradient in two ways: the 

gradient value is not too large when the difference between 

the prediction frame and the ground truth is too large.  The 

gradient value is small enough when the difference between 

the prediction frame and ground truth is small. 

𝐿
smooth𝐿1 = {

1

𝑁
∑ ||𝑥, 𝑦||1

𝑁
𝑖=1

1

𝑁
∑ ||𝑥, 𝑦||2

𝑁
𝑖=1

            (3) 

The total loss function of the network is: 

𝐿
total

= 𝐿𝑆𝑆𝐼𝑀 + 𝐿𝑠𝑚𝑜𝑜𝑡ℎ𝐿1           (4) 

3. Experiments and Results Analysis 

3.1. Experimental details 

The experiment was conducted on Ubuntu 20.04.3 with a 

NVIDIA RTX 3090 24GB×2. Pytorch-gpu is used for the 

deep learning architecture, epochs are 150, and batch size is 

8. The models are all optimized using the ADAM optimizer 

with an initial learning rate of 0.0005. 

The LOL dataset is the training set for the experiment. A 

dataset containing 1000 low-light and normal exposure pairs 

with an image size of 400x400x3, and 985 pairs are used in 

training and 15 pairs for validation is included. It consists of 

the evaluation sets of LOL and SICE, as well as the data sets 

of Exrdark[16], DICM[17], VV[18], NPE[19], MEF[20] and 

LIME[6].  

Peak Signal-to-Noise Ratio (PSNR), Structural Similarity 

(SSIM), Natural Image Quality Evaluation (NIQE) and Root 

Mean Square Error (RMSE) are used as objective evaluation 

indicators. PSNR and SSIM values are higher when the 

enhanced image is better. An enhanced image is of higher 

quality if the NIQE and RMSE values are low. 

 There are 8 mainstream low-light enhancement 

algorithms in the current research field: LIME, BIMEF, 

RetinexNet, SDD, CSDGAN, RUAS, KinD++, 

EnlightenGAN. Images are tested using models or source 

code from the original literature. 

3.2. Comparison Experiment 

The results of the reference LOL and SICE datasets are 

shown in Table 1. In both PSNR and SSIM metrics, this 

algorithm achieves the best results, and in LOE metrics, it 

holds the second place, indicating that this algorithm 

significantly enhances the quality of low-light images. Table 

2 shows the results of NIQE metrics on 6 datasets. Despite 

not achieving the optimal results on all datasets, the algorithm 

in this paper has a high advantage on average and does 

achieve the optimal results on most datasets. 

Figure 2 is a comparison chart of the enhancement effect 

on 4 datasets, and it can be seen that the image color saturation 

after the enhancement of this method is high, there is no 

obvious noise and artifacts, and it has a rich detail texture. 

MBLLEN and SDD methods have the phenomenon of poor 

brightness, low saturation and blur of the enhanced image. 

The image enhanced by the Retinex-Net method has obvious 

color distortion problems, resulting in blurring of some areas 

in the graph and amplifying the noise in the image; The 

EnlightenGAN method has a good effect of enhancement. 

but there will be some problems of poor exposure, and the 

halo problem is more serious; Kinder++ has improved on 

color distortion, but there is still a problem that the enhanced 

image is not realistic and will amplify the noise in local areas. 

In summary, our model has achieved excellent performance 

in the comprehensive performance of multi-objective 

enhancement tasks on the test data set, which can adaptively 

increase the global brightness and avoid overexposure; Can 

correct the color distortion in low-light images, enhance the 



 

27 

result of rich and full color, and the visual effect is better; Has 

generalization enhanced for low-light images of real scenes. 

 

Table 1. Comparison of indicators of different methods on reference datasets (Where means higher value is better, means 

lower value is better, red means the best result, blue means the second-best result) 

Dataset LOL SICE 

Metric PSNR↑ SSIM↑ RMSE↓ PSNR↑ SSIM↑ RMSE↓ 

LIME 13.374 0.508 37.2 15.798 0.625 45.6 

BIMEF 13.921 0.728 40.3 13.544 0.533 56.1 

RetinexNet 17.612 0.653 34.6 15.784 0.616 41.3 

SDD 13.352 0.640 60.4 13.388 0.533 59.5 

CSDGAN 9.111 0.331 90.1 10.464 0.381 81.9 

RUAS 16.532 0.527 51.2 10.578 0.396 43.6 

EnlightenGAN 17.546 0.664 38.3 14.768 0.611 50.3 

KinD++ 17.789 0.766 34.8 14.671 0.513 49.2 

Ours 21.365 0.793 20.9 16.453 0.664 42.1 

Table 2. Comparison of NIQE indicators of different methods (↑ means higher value is better, means lower value is better, red means the 
best result, blue means the second-best result) 

Metric Exrdark DICM LIME VV NPE MEF AVERAGE 

LIME 3.855 3.459 4.077 2.920 3.595 3.025 3.489 

BIMEF 3.829 3.541 3.861 3.239 3.607 3.333 3.568 

RetinexNet 4.451 4.193 4.825 2.983 4.364 3.982 4.133 

SDD 4.120 3.937 4.397 3.281 3.701 3.857 3.882 

CSDGAN 5.591 5.782 5.730 4.730 5.179 7.338 5.725 

RUAS 4.404 5.204 5.305 5.326 5.380 5.617 5.206 

EnlightenGAN 3.691 3.563 3.664 4.237 3.731 3.221 3.684 

KinD++ 4.110 3.804 4.722 2.806 3.593 3.737 3.795 

Ours 3.592 3.512 3.763 3.115 3.359 3.391 3.455 

 
Figure 2. Visual comparison of different methods. 

 

4. Conclusion 

Aiming at the problem of low light enhancement, this paper 

proposes a new multi-scale low-light image enhancement 

network with regularized contrast. Two positive and negative 

samples are used to guide the network in learning. Through 

experimental verification, the proposed algorithm is far 

superior to most existing low-light image enhancement 

algorithms, which enhances the quality of low-light images 

and provides data for subsequent high-order picture 

processing tasks. In the next work, we will consider 

optimizing the computing performance of the network so that 

it can be combined with more realistic scene information, and 

try to use the model for low-light video processing.   



 

28 

References 

[1] Kim, Y.T., “Contrast enhancement using brightness preserving 
bihistogram equalization,” IEEE transactions on Consumer 
Electronics, 1–8(1997).  

[2] Land, E. and McCann, J., “Lightness and retinex theory,” Josa, 
vol. 61, no. 1, pp. 1–11, (1971). 

[3] Guo, X., Li, Y. and Ling, H., “Lime: Low-light image 
enhancement via illumination map estimation,” IEEE 
Transactions on image processing, vol. 26, no. 2, pp. 982–
993(2016). 

[4] Zhao, A., Dong, J., Li, J., Qi, L. and Zhou, H., "Associated 
Spatio-Temporal Capsule Network for Gait Recognition," in 
IEEE Transactions on Multimedia, doi: 
10.1109/TMM.3060280(2021) . 

[5] Wang, Y., Li, J., Zhao, A., Lv, Z., and Lu, G., “Temporal 
Attention-Based Graph Convolution Network for Taxi 
Demand Prediction in Functional Areas,” In International 
Conference on Wireless Algorithms, Systems, and 
Applications, 203-214 Springer (2021). 

[6] Zhao, A., Li, J., Dong, J., Qi, L., Zhang, Q., Li, N., Wang, X. 
and Zhou, H., "Multimodal Gait Recognition for 
Neurodegenerative Diseases," in IEEE Transactions on 
Cybernetics, doi: 10.1109/TCYB.3056104(2021).  

[7] GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature 
hierarchies for accurate object detection and Semantic 
segmentation[J]. Proceedings of 2014 IEEE Conference on 
Computer Vision and Pattern Recognition. Columbus: IEEE, 
2014: 580-587. 

[8] GIRSHICK  R.  Fast  R-CNN  [C]//Proceedings  of  2015  
TEEE  International  Conference  on Computer Vision. 
Santiago: IEEE, 2015: 1440-1448 Hao, S., Han, X., Guo, Y., 
Xu, X. and Wang, M., “Low-light image enhancement with 
semi-decoupled decomposition,” IEEE transactions on 
multimedia, vol. 22, no. 12, pp. 3025–3038(2020). 

[9] REN S Q, HE K M, GIRSHICK R, et al. Faster R-CNN; 
towards real-time object detection with  region  proposal  
network[J].  IEEE  Transactions  on  Pattern  Analysis  
and  Machine Intelligence, 2017, 39(6): 1137-1149. 

[10] Redmon J, Divvala S, Girshick R, et al. You Only Look Once: 
Unified, Real-Time Object Detection[C]//Proceedings of the 
IEEE conference on computer vision and pattern recognition. 
2016: 779-788. 

[11] LIU W, ANGUELOV D, ERHAN D, et al. SSD: Single shot 
multibox detector [C]//European conference on computer 
vision. Springer, Cham, 2016: 21-37. 

[12] Ma, L., Liu, R., Zhang, J., Fan, X. and Luo, Z., “Learning deep 
context-sensitive decomposition for low-light image 
enhancement,” IEEE Transactions on Neural Networks and 
Learning Systems (2021). 

[13] Sun, Y., Zang, Y. and Liu, S., “Image Super-Resolution Using 
Supervised Multi-scale Feature Extraction Network,” 
Multimedia Tools and Applications,80(2): 1995－2008(2020). 

[14] Yuanjie Shao, Lerenhan Li, Wenqi Ren, Changxin Gao, and 
Nong Sang. Domain adaptation for image dehazing. In CVPR, 
2020. 1, 2, 4. 

[15] L. Li, Y. Dong, W. Ren, J. Pan, C. Gao, N. Sang, and M. Yang. 
Semi-supervised image dehazing. TIP, 29:2766–2779, 2020. 4. 

[16] Lee, C., Lee, C. and Kim, C., “Contrast enhancement based on 
layered difference representation of 2d histograms,” IEEE 
transactions on image processing, vol. 22, no. 12, pp. 5372–
5384(2013). 

[17] Cai, J., Gu, S. and Zhang, L., “Learning a deep single image 
contrast enhancer from multi-exposure images,” IEEE 
Transactions on Image Processing, vol. 27, no. 4, pp. 2049–
2062(2018). 

[18] Wang, S., Zheng, J., Hu, H. and Li, B., “Naturalness preserved 
enhancement algorithm for non-uniform illumination images,” 
IEEE transactions on image processing, vol. 22, no. 9, pp. 
3538–3548 (2013). 

[19] Ma, K., Zeng, K. and Wang, Z., “Perceptual quality assessment 
for multi-exposure image fusion,” IEEE Transactions on Image 
Processing, vol. 24, no. 11, pp. 3345–3356(2015). 

[20] Mittal, A., Sundararajan, R. and Bovik, A., “Making a 
“completely blind” image quality analyzer,” IEEE Signal 
processing letters, vol. 20, no. 3, pp. 209–212(2012).