Applying Genetic Algorithm and Image Quality Assessment for

Reproducible Processing of Low-light Images

Olivier Parisot and Thomas Tamisier

Luxembourg Institute of Science and Technology (LIST),

5 Avenue des Hauts-Fourneaux, 4362 Esch-sur-Alzette, Luxembourg

Keywords:

Low-light Images, Image Quality Assessment, Genetic Algorithms.

Abstract:

Reproducible images preprocessing is fundamental in computer vision, whether to fairly compare process

algorithms or to prepare new images corpus. In this paper, we propose an approach based on genetic algorithm

combined to Image Quality Assessment methods to obtain a reproducible sequence of transformations for

improving low-light images. Preliminary tests have been performed on state-of-the-art benchmarks.

1 INTRODUCTION

Images captured in poor lighting conditions often ex-

hibit characteristics such as low brightness, low con-

trast, narrow gray scale, color distortion, and high

noise - making them difﬁcult for the human eyes to

view details (Wang et al., 2020). Improvement of the

quality of such images is a popular research area in

computer vision.

In general, applying appropriate transformation

to improve given image requires powerfull tools and

strong expertise (Chaudhary et al., 2018). For in-

stance, a regular user of dedicated software like Gimp

or Photoshop process images by incrementally creat-

ing/modifying/merging layers until the result is sat-

isfying. In order to automate as much as possible

this workﬂow, two elements are essential. On the one

hand, it is important to use speciﬁc metrics to guide

the process: in this regard, Image Quality Assessment

aims at estimating the quality of an image in a way

that corresponds to a human subjective scoring of the

same image (Zhai and Min, 2020). On the other hand,

new techniques are constantly proposed in the liter-

ature to enhance images (Parekh et al., 2021); nev-

ertheles, most of them are based on Deep Learning

techniques that produce effective results – the effec-

tive transformation is then difﬁcult to interpret or re-

produce by another method (Buhrmester et al., 2021).

This is particularly the case for low-light images, as

shown by a recent survey presenting the recent works

(Li et al., 2021).

However, in the context of academic research or

industrial innovation, it is increasingly required to

guarantee the reproducibility of experiments by keep-

ing trace of the transformations performed on the im-

ages (Berg, 2018). As an example, a recent paper has

shown that an important proportion of research works

lacks of transparency regarding image handling and it

may compromise the interpretation of the leading re-

sults (Miura and Nørrelykke, 2021). We can make an

Figure 1: High resolution photography of a telescope cap-

tured by the author during the night time by using a smart-

phone. The picture was not processed by an additional soft-

ware – only a minimal treatment was applied by the smart-

phone ﬁrmware.

Parisot, O. and Tamisier, T.

Applying Genetic Algorithm and Image Quality Assessment for Reproducible Processing of Low-light Images.

DOI: 10.5220/0011082400003209

In Proceedings of the 2nd International Conference on Image Processing and Vision Engineering (IMPROVE 2022), pages 189-194

ISBN: 978-989-758-563-0; ISSN: 2795-4943

189

analogy with Machine Learning: data preprocessing

should be transparent in order to lead to meaningful

and trustable predictive models (Zelaya, 2019).

In this paper, we propose an approach based on a

genetic algorithm to obtain a reproducible improve-

ment of low-light images quality by relying on trans-

formations monitored by Image Quality Assessment

methods.

The rest of this article is organized as follows.

Firstly, related works about Image Quality Assess-

ment and image quality improvement of low-light im-

ages are brieﬂy presented (Section 2). Then, our ap-

proach to improve the quality of low-light images is

described (Section 3). Finally, a concrete implemen-

tation is detailed (Section 4), the results of prelim-

inary experiments are discussed (Section 5) and we

conclude by opening some perspectives (Section 6).

2 RELATED WORKS

2.1 Image Quality Assesment

Numerous Image Quality Assessment approaches

were developed in recent years and an exhaustive list

was already compiled (Zhai and Min, 2020). They

are is widely used in benchmarks to compare the ef-

ﬁciency of image processing algorithms (Li et al.,

2018).

We can distinguish two main types of tech-

niques: Full-reference (FR) and Reduced-reference

(RR) methods are based on a referential of images

(raw/distorted) while No-reference (NR) and Blind

methods intend to estimate single image quality (Liu

et al., 2019). In this paper, we prefer to focus on NR

and Blind approaches because because in practice it

is very often difﬁcult to obtain both raw and corrected

images. Among them, we can mention:

• Classical methods like BRISQUE (Blind/Refer-

enceless Image Spatial Quality Evaluator): a

score between 0 and 100 is produced (0 for good

quality image, 100 for poor quality) (Mittal et al.,

2012).

• Recent Deep Learning methods like NIMA (Neu-

ral Image Assessment) – a set of Convolutional

Neural Networks to estimate the aesthetic and

technical quality of images: a score between 0 and

10 is produced (0 for poor quality, 10 for good

quality) (Talebi and Milanfar, 2018).

• Dedicated techniques for low-light images like

NLIEE (No-reference Low-light Image Enhance-

ment Evaluation) (Zhang et al., 2021): the leading

quality score represents various aspects like light,

color comparison, noise and structure.

2.2 Genetic Algorithm for Images

Processing

Nature Inspired Optimization is a family of problem-

solving approaches derived from natural processes.

Among them, the most popular include genetic al-

gorithms and particle swarm optimization (Li et al.,

2020). These approaches are increasly applied in im-

age processing for various tasks such as blur and noise

reduction, restoration and segmentation (Dhal et al.,

2019; Ramson et al., 2019). In particular, (Parisot

and Tamisier, 2021) process images with a Nature In-

spired Optimization Algorithm.

To the best of our knowledge, there are no much

contributions about the reproducible transformations

of low-light images by applying genetic algorithm

guided by Image Quality Assessment techniques.

3 APPROACH

The cornerstone of our approach is deﬁned as follows:

• An initial low-light image.

• A sequence of speciﬁc transformations applied on

the initial image (examples: brighten, enhance,

dehaze, adjust histogram, deblur, total variation

denoise, etc.).

• A quality score evaluated by using a method S.

This step is critical and drives the algorithm (qual-

ity serves here as the ﬁtness of the solution, in the

terminology used for evolutionary algorithms).

For a given low-light input image (I), by considering a

quality evaluation method (S) and a maximum count

of epochs (E), the following genetic algorithm com-

putes the transformations sequences leading to an im-

age with a better quality:

• A population is generated with P images: each

image is a clone of the initial image I on which

a random transformation has been applied or not.

In fact, to ensure that the algorithm does not lead

to a lower-quality image, it is important to keep

at least one unmodiﬁed clone of the initial image

in the population: at worst, it will remain the best

solution.

• During E epochs:

– The current best image or an other randomly

selected image is cloned, and then a random

transformation is applied: the newly created

IMPROVE 2022 - 2nd International Conference on Image Processing and Vision Engineering

190

image is evaluated with S and added into the

population.

– An other image is randomly selected in the pop-

ulation and is stacked with the initial image

(with a random weight): the newly created el-

ement is evaluated with S and added into the

population.

– According to the evaluation with S of the im-

ages present in the population, the worst images

are selected and then removed from the popu-

lation (to always keep P images in the popula-

tion).

• The ﬁnal result is the image of the consolidated

population having the best quality estimation. The

algorithm output is then an sequence of transfor-

mations that leads to an amelioration of the Image

Quality Assessment.

The quality score resulting from the method (S) is

evaluated by using both a Image Quality Assessment

method and a brightness estimation. The quality esti-

mator will be able to evaluate the global quality of the

image while an explicit estimation of brightness may

help to give a better score to brighter images as they

tend to exhibit more details. As a result, we propose

a quality score method (S) deﬁned as follows:

• the quality score is the result of a selected Image

Quality Assessment method.

• if the brightness of the image being evaluated is

lower than that of the reference image, then a

malus is applied to the score.

• Conversely, if the brightness of the image being

evaluated is higher than that of the reference im-

age, then a bonus is applied to the score.

To prevent the image from deviating too much from

the original one, we have added a test comparing the

similarity between the produced image and the ini-

tial image: if the similarity is too low (i.e. lower

than a predeﬁned threshold T), then the image score is

strongly penalised and the last transformation is there-

fore not retained. The test is based here on the Struc-

tural Similarity Index (SSIM): in practice, the value is

close to 1 when the two images are similar while the

value is close to 0 when the images are really differ-

ent.

4 PROTOTYPE

The algorithm has been implemented into a Python

prototype. Various well-known open-source packages

have been integrated. Images loading and transfor-

mations are realized with various dedicated packages

like openCV

and scikit-images

. BRISQUE score

is computed through the image-quality package

and

NIMA scores are provided by a Tensorﬂow imple-

mentation

By using these packages, these image transforma-

tions can be applied:

• Blurring and deblurring.

• Denoising/restoration: total variation, non local

means, wavelets, bilateral, Noise2Noise (Lehti-

nen et al., 2018).

• Contrast adjustment / Histogram optimization by

using CLAHE (Contrast Limited Adaptive his-

togram equalization, (Zuiderveld, 1994)).

• Background estimation and processing (Guo and

Wan, 2018).

• Dehazing via Deep Learning methods like Cycle-

Dehaze (Engin et al., 2018).

• Morphological transformations (like erode and di-

late) (Sreedhar and Panlal, 2012).

Moreover, the brightness was evaluated by a method

proposed by (Rex Finley, 2006).

The prototype was tested on a computing infras-

tructure with the following hardware conﬁguration:

40 cores and 128 GB RAM (Intel(R) Xeon(R) Sil-

ver 4210 CPU @ 2.20GHz) and NVIDIA Tesla V100-

PCIE-32GB. The CUDA

et NUMBA

frameworks

have been used to optimize the usage of the hardware

(CPUs and GPUs).

5 FIRST EXPERIMENTS

The prototype was executed on low-light benchmarks,

i.e. with images coming fron the LOL dataset (Wei

et al., 2018) and the VIP-lowLight dataset (Chung and

Wong, 2016), as shown in Figure 2 and in Table 1.

The presented method can thus be seamlessly in-

serted into any image processing workﬂow; not only

is it possible to reproduce the image processing se-

quence, but it also allows to modify it afterwards

if needed (for manual adjustments according to the

speciﬁcities of the images – such as additional denois-

ing).

Moreover, the ﬁrst experiments show that the re-

sults obtained on the benchmarks are globally satis-

factory. Table 1 and Table 2 have been obtained with

https://pypi.org/project/opencv-python/

https://pypi.org/project/scikit-image/

https://pypi.org/project/image-quality/

https://github.com/idealo/image-quality-assessment

https://developer.nvidia.com/cuda-zone

http://numba.pydata.org/

Applying Genetic Algorithm and Image Quality Assessment for Reproducible Processing of Low-light Images

191

(a) The original raw image.

(b) The processed image.

Figure 2: Raw image coming from the LOL dataset (2a), and the second one was processed with our algorithm (2b). The

original raw image has the following characteristics: BRISQUE=25.0677, noise variance=2.779.The following sequence has

been computed: adjust gamma (sigma=1), sum with 0.7x the original image, sum with 0.2x the original image, increase con-

trast, CLAHE (clipLimit=1). The processed image has the following characteristics: BRISQUE=6.2736, noise variance=5.54.

(a) The original raw image.

(b) The processed image.

Figure 3: An other raw image coming from the LOL dataset (3a), and the second one was processed with our algorithm (3b).

The original raw image has the following characteristics: BRISQUE=21.8989, noise-variance=2.798 The following sequence

has been computed: blur (sigma=0.05), enhance (factor=1.05), CLAHE (clipLimit=2), sum with 0.3x the original image,

enhance (factor=0.95), CLAHE (clipLimit=1). The processed image has the following characteristics: BRISQUE=20.1891,

noise-variance=6.523.

the following hyperparameters: BRISQUE as target-

ted Image Quality Assessment score combined with

brightness control, an initial population of 20 images,

50 maximum epochs and 0.25 as minimum similarity.

According to signiﬁcant runs, this setting offers the

best tradeoff between quality improvement and exe-

cution time. BRISQUE score has been computed af-

terwards to check the quality of the algorithm inputs

/ outputs and Noise Variance (Immerkaer, 1996) has

been esimated to highlight the level of noises in the

benchmark.

Finally, a word on performances: the time needed

for the experiments was reasonnable on the infrastruc-

ture described above (from a few seconds to several

dozen seconds per image – depending of the images

shapes). During our preliminary tests, we have ran

the algorithm on small (Figure 2) and large images

(Figure 1) – and the computation time was not the

same: the image transformation operations obviously

took more time on high resolution images. In prac-

tice, a tradeoff between algorithm efﬁciency and ex-

ecution time is required, and it may be controlled by

the genetic algorithm settings (epochs count, popula-

tion size, etc.). An other trick consists in using min-

imzed version of raw images during the genetic al-

gorithm execution (let say by reducing the size by a

factor of let two): once the sequence is calculated, it

can be further applied to the original image. The qual-

ity evaluation will be less precise, but it will greatly

accelerate the execution of the algorithm.

IMPROVE 2022 - 2nd International Conference on Image Processing and Vision Engineering

192

Table 1: Experiments on the VIP-LowLight benchmark: the

(average,min,max) values are listed for each metric (before

the algorithm execution).

Raw images

BRISQUE (22.8029, 21.0121, 26.0876)

NIMA-aesthetic (4.3338, 4.0052, 4.6494)

NIMA-technical (4.3953, 4.1569, 4.9886)

Noise variance (5.7839, 1.701, 10.696)

Table 2: Experiments on the VIP-LowLight benchmark: the

(average,min,max) values are listed for each metric (after

the algorithm execution).

Processed images

BRISQUE (2.1749, 0.3492, 13.7365)

NIMA-aesthetic (4.7018, 3.9337, 5.8975)

NIMA-technical (4.7208, 4.2394, 5.0744)

Noise variance (11.855, 2.337, 71.425)

6 CONCLUSION

This paper presented an approach based on a genetic

algorithm to improve the quality of a given low-light

images from a reproducible sequence of transforma-

tions. A prototype based on Image Quality Assess-

ment methods was implemented and tested on various

state-of-the-art low-light images databases.

Thanks to academic and operational partners, we

will set-up real-world use-cases to validate the ap-

proach. In parallel, we will improve the prototype

by automatically generating the Python source code

to transform the image as provided by Automated

Machine Learning platforms for predictive models.

Finally, we will work to improve execution perfor-

mance by distributing calculations via frameworks

like Spark because the Map/Reduce concept may

drastically speed-up genetic algorithms execution.

ACKNOWLEDGMENTS

This work was carried our during the MILAN project

(MachIne Learning for AstroNomy) – funded by the

Luxembourg National Research Fund. The tests were

realized on the LIST Artiﬁcial Intelligence and Data

Analytics platform (LIST AIDA). Special thanks to

Raynald Jadoul and Jean-Franc¸ois Merche for their

support.

REFERENCES

Berg, J. (2018). Progress on reproducibility.

Buhrmester, V., M

unch, D., and Arens, M. (2021). Analysis

of explainers of black box deep neural networks for

computer vision: A survey. Machine Learning and

Knowledge Extraction, 3(4):966–989.

Chaudhary, P., Shaw, K., and Mallick, P. K. (2018). A sur-

vey on image enhancement techniques using aesthetic

community. In International Conference on Intel-

ligent Computing and Applications, pages 585–596.

Springer.

Chung, A. G. and Wong, A. (2016). Noise suppression and

contrast enhancement via bayesian residual transform

(brt) in low-light conditions. Journal of Computa-

tional Vision and Imaging Systems, 2(1).

Dhal, K. G., Ray, S., Das, A., and Das, S. (2019).

A survey on nature-inspired optimization algorithms

and their application in image enhancement domain.

Archives of Computational Methods in Engineering,

26(5):1607–1638.

Engin, D., Genc¸, A., and Ekenel, H. K. (2018). Cycle-

dehaze: Enhanced cyclegan for single image dehaz-

ing. In The IEEE Conference on Computer Vision and

Pattern Recognition (CVPR) Workshops.

Guo, Y. and Wan, Y. (2018). Image enhancement algo-

rithm based on background enhancement coefﬁcient.

In 2018 10th International Conference on Communi-

cations, Circuits and Systems (ICCCAS), pages 413–

417.

Immerkaer, J. (1996). Fast noise variance estimation. Com-

puter vision and image understanding, 64(2):300–

302.

Lehtinen, J., Munkberg, J., Hasselgren, J., Laine, S., Kar-

ras, T., Aittala, M., and Aila, T. (2018). Noise2noise:

Learning image restoration without clean data. arXiv

preprint arXiv:1803.04189.

Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., and

Wang, Z. (2018). Benchmarking single-image dehaz-

ing and beyond. IEEE Trans. on Image Processing,

28(1).

Li, C., Guo, C., Han, L.-H., Jiang, J., Cheng, M.-M., Gu, J.,

and Loy, C. C. (2021). Low-light image and video en-

hancement using deep learning: a survey. IEEE Trans-

actions on Pattern Analysis & Machine Intelligence,

(01):1–1.

Li, H., Liu, X., Huang, Z., Zeng, C., Zou, P., Chu, Z.,

and Yi, J. (2020). Newly emerging nature-inspired

optimization-algorithm review, uniﬁed framework,

evaluation, and behavioural parameter optimization.

IEEE Access, 8:72620–72649.

Liu, Y.-H., Yang, K.-F., and Yan, H.-M. (2019). No-

reference image quality assessment method based on

visual parameters. Journal of Electronic Science and

Tech., 17(2).

Mittal, A., Moorthy, A. K., and Bovik, A. C. (2012).

No-reference image quality assessment in the spa-

tial domain. IEEE Trans. on image processing,

21(12):4695–4708.

Miura, K. and Nørrelykke, S. F. (2021). Reproducible

image handling and analysis. The EMBO journal,

40(3):e105889.

Applying Genetic Algorithm and Image Quality Assessment for Reproducible Processing of Low-light Images

193

Parekh, J., Turakhia, P., Bhinderwala, H., and Dhage, S. N.

(2021). A survey of image enhancement and object

detection methods. Advances in Computer, Commu-

nication and Computational Sciences, pages 1035–

1047.

Parisot, O. and Tamisier, T. (2021). Reproducible improve-

ment of images quality through nature inspired opti-

misation. In International Conference on Cooperative

Design, Visualization and Engineering, pages 335–

341. Springer.

Ramson, S. J., Raju, K. L., Vishnu, S., and Anagnostopou-

los, T. (2019). Nature inspired optimization tech-

niques for image processing—a short review. NIO

Techniques for Image Processing Applications, pages

113–145.

Rex Finley, D. (2006). Hsp color model — alternative to

hsv (hsb) and hsl.

Sreedhar, K. and Panlal, B. (2012). Enhancement of images

using morphological transformation. arXiv preprint

arXiv:1203.2514.

Talebi, H. and Milanfar, P. (2018). Nima: Neural im-

age assessment. IEEE Trans. on Image Processing,

27(8):3998–4011.

Wang, W., Wu, X., Yuan, X., and Gao, Z. (2020). An

experiment-based review of low-light image enhance-

ment methods. IEEE Access, 8:87884–87917.

Wei, C., Wang, W., Yang, W., and Liu, J. (2018). Deep

retinex decomposition for low-light enhancement.

arXiv preprint arXiv:1808.04560.

Zelaya, C. V. G. (2019). Towards explaining the effects

of data preprocessing on machine learning. In 2019

IEEE 35th ICDE, pages 2086–2090. IEEE.

Zhai, G. and Min, X. (2020). Perceptual image quality as-

sessment: a survey. Science China Information Sci-

ences, 63:1–52.

Zhang, Z., Sun, W., Min, X., Zhu, W., Wang, T., Lu, W., and

Zhai, G. (2021). A no-reference evaluation metric for

low-light image enhancement. In 2021 IEEE Interna-

tional Conference on Multimedia and Expo (ICME),

pages 1–6. IEEE.

Zuiderveld, K. (1994). Contrast Limited Adaptive His-

togram Equalization, page 474–485. Academic Press

Professional, Inc., USA.

IMPROVE 2022 - 2nd International Conference on Image Processing and Vision Engineering

194