Low Light Image Enhancement Using Zero-DCE algorithm

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 04 | Apr 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 200
Low Light Image Enhancement Using Zero-DCE algorithm
Nikhil Patil1, Riddhesh Pingle2, Kalyani Sarak3
[1][2][3]Student, Department of Information Technology, Atharva College of Engineering, Mumbai
[4]Professor Deepali Maste, Department of Information Technology, Atharva College of Engineering
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - In modern world image enhancement is one of
the complex task. The proposed Low Light image
enhancement using Zero DCE (Zero Reference Deep Curve
Estimation) algorithm improves the quality, color
correction and accuracy of the image. It is useful in many of
the applications such as military applications, object
detection, recognition, and also in 3D reconstruction of
images. When image is captured in low light objects are not
clearly visible, lot of noise and image disturbances present
because of that image is not clearly visible. Normally images
contains light over specific range of wavelength
corresponding to the visible portion of the spectrum.By
using Zero DCE(Zero Reference Deep Curve Estimation) we
improved the image quality and accuracy via DCE-Net(
Deep Curve Estimate Network) and LE(light Enhancement )
curve.
1. INTRODUCTION
Images plays very vital role in modern world. Detecting
the object in low light image is very challenging task. In
today’s world low light image enhancement is highly
demanded. It needs to improve the visual appearance of
the image to provide better transform for future image
processing. Low light images contains noise, low contrast
which needs to be enhance and produce accurate, denoise
image. At some movements of life we can’t click images
again and again like our childhood images. We need
cleared images that can be done by using Zero DCE.
Various image enhancement algorithms are proposed,
they focuses on contrast enhancement. However the Zero
DCE algorithm focuses on uniform enhancement. We
present the framework of Zero-DCE (Zero Reference Deep
Curve Estimation). The algorithm contains DCE-Net, LE
curve and non-reference loss function. DCE-Net algorithm
estimates the best fitting LE Curves onto an given input
image. Framework performs mapping of input version’s
RGB channel pixels by applying curves iteratively for
obtaining final enhanced version. DCE-Net finds the best
fitting curve parameters between input image and output
image. It improves the pixel quality of the dark image.
Then the next stage of the image is LE curve. We
iteratively apply LE curve which contains parameters α
and number of iterations n. Then next stage of image is
non reference loss function that evaluate the quality of the
enhance image. This algorithm is also works with DSLR
images and dark video.
1.1 Motivation
In today’s hustle bustle life the technologies main goal is to
make the things more clear, accurate, less time consuming
and easy to understand. As we know that images plays
very precious role in our life like by using images we can
represent whole life journey in a few minutes. There is
some condition where images are captured in low light
circumstances because of which objects inside images are
not clearly visible. So, to overcome this problem we
proposed a system for low light image enhancement using
Zero DCE which take dark images as an input and
producing enhanced images as an output.
1.2 AIM and Objective
The main aim of this project is to accept the low light
image as input and produce the enhanced image as output
by using Zero DCE (Zero Reference Deep Curve
Estimation) algorithm.
Following are the objectives:
1 .To be used in many real application like automated
driving.
2. To be used for task like for classification, segmentation,
recognition, scene understanding and also for 3D
reconstruction of images.
1.3 Basic Concept
Our solution is to transfer the dark image as enhance
image via Zero DCE algorithm. The algorithm focuses on
the color correction. If the input image has some noise ,
Zero DCE result will also has noise, hence we improve the
result by doing denoising .Zero DCE focuses on three steps
such as color correction, Denoising and retrain the model.
2. Background study:
The basic idea of this project came from a background
study of [3] which formulates light enhancement as a task
of image-specific curve estimation with a deep network
and “A CNN-Based Method to Enhance Low-Light Remote-
Sensing Images” written by Linshu Hu , Mengjiao Qin ,
Feng Zhang , Zhenhong Du and Renyi Liu[10] they used
SRCNN(Super Resolution CNN) , Remote Sensing images
are enhanced on basis of new architecture using
SRCNN(Super Resolution CNN)[10]

Li Tao , Chuang Zhu, Jiawen Song, Tao Lu, Huizhu Jia
implemented Low Light image enhancement using CNN
and channel priority.Using CNN framework to denoise an
low light image. Then on basis of atmosphere scattering
model produced an simple enhanced image prior with high
contrast. Denoising done using AWGN(Additive White
Gaussian Noise) method. On basis of VDSR(Super
resolution CNN) and DnCNN(denoising network CNN) they
proposed a denoising CNN to remove image noise. Then
the noise free image is converted to enhanced image. In
this paper, a joint effective method is proposed by
combining denoising and contrast enhancement for low-
light images.[2]
3. Block Diagram:
Input Image: Noisy Image which will be denoised using
STN(Structure Texture Noise Decomposition).
Zero DCE: Zero Reference Deep Curve Estimation
algorithm used to enhance low light image using a zero
reference image.
Convolution Layer: The original input image is converted
into greyscale matrix. After transformation of image into
big matrix we multiplied this big matrix with predefined
small matrix of sharpen convolution layer. The new
version of image we get has clearly visible edges.[2]
CNN Model:While constructing a Neural Network, at the
initial stage, we initialize weights with some random values
or any variable for that fact. It’s not necessary that
whatever values of the weights we have selected need to be
correct or it fits the model best.
Color Restoration:In this step greyscale images restore
their colour using RGB space values. Output Image: The
result of above steps is an Enhanced Image eg: Fig1.1
Fig -1.1
4. Algorithms:
4.1.1:
Light Enhancement Curve (LE-Curve):
The curve adjustments methods used in photo editing
softwares inspired us to implement the same design such
that , a LE curve which can map a low light image to its
enhanced version.
Self adaptive curve parameters depends on input image.
Three objectives of the design:
A) The pixel value of each in the enhanced version should
be within range [0,1] , leads to avoiding information loss at
time of overflow truncation.
B) Monotonous curve to preserve contrast of
neighbouring pixels.
C) Form of curve should be differentiable before gradient
propagation.
Below is an illustration of the mechanism:
Eq 1 --> LE(I(x); α) = I(x) + α I(x)(1 − I(x)),
Here, x is denoted as pixel coordinates
LE(I(x); α) denotes enhanced version of input.
I(α) is the trainable curve parameter , must be within
[-1, 1].
Exposure level in the input is controlled by I(α).
The LE- Curve is applicable for not only enhancing the
darker regions but also diminishing over-exposure effects.
4.1.1.1:
Higher Order Curves:
Equation 1 can be modified as following to iteratively
enable more versatile modifications to the input to
overcome low light conditions .Specifically,
LE n(x) = LE n−1(x) + α n LE n−1 (x)(1 − LE n−1(x)), (2)
4.1.1.2
Pixel-Wise Curve:
A limitation of higher order curve is that it produces an
over/under enhanced local regions in input image. To
address the issue we formulate α as pixel wise parameter,
each pixel value will have a corresponding best-fitting
curve to modify its dynamic range.

Hence, Eq.(2)
LEn(x) = LEn−1(x)+An(x)LEn−1(x)(1−LEn−1(x)), (3)
Here, A is a parameter map with the same size as the
given image. Here, we assume that pixels in a local region
have the same intensity (also the same adjustment
curves), and thus the neighboring pixels in the output
result still preserve the monotonous relations. In this way,
the pixel-wise higher-order curves also comply with three
objectives.[3]
4.1.2
ZERO DCE:
A DCE net algorithm maps the input image to its best
fitting curve. Zero DCE takes an low light image as an input
and produces an enhanced version. Deploying a plain CNN
of 7 convolution layers with a symmetrical concatenation.
The DCE net architecture has 32 convolution kernels of
3x3 size and 1 stride. Discarding the down sampling and
normalizing training batch. Training data consists of
79,416 images of 256x256x3 size.
4.1.2.1
DCE-Net ‘s Non-Reference Loss Functions :
To enable zero-reference learning in DCE-Net, we provide
diffrentiable non-reference losses that allows to evaluate
the quality of enhanced version. The following four types
of losses are adopted to train our DCE-Net[3]. Spatial
Consistency Loss. The spatial consistency loss Lspa
encourages spatial coherence of the enhanced image
through preserving the difference of neighboring regions
between the input image and its enhanced version: Lspa = 1
K K i=1 j∈Ω(i) (|(Yi − Yj )|−|(Ii − Ij )|) 2, We denote Y and I
as average of intensity of exposure in the enhanced
version and input .We empirically set the size of the local
region to 4×4.
Input
Zero-DCE
w/o LSPA
W/0 Lexp

W/0 LCOL
w/0 LtvA
Ablation study of the contribution of each loss (spatial
consistency loss Lspa, exposure control loss Lexp, color
constancy loss Lcol, illumination smoothness loss LtvA).
4.1.3
CNN FOR DENOISING:
Most commonly, noise signal in images is assumed to be
independent, identically distributed. Thus, denoise
methods usually focus attention on the problem of
attenuating additive white Gaussian noise (AWGN). These
methods assume that the standard deviation σ of the
AWGN has been accurately estimated. Many methods do
estimate the variance accurately in most cases but still
have non-negligible errors in some special situations.[2]
4.1.4:
Illuminant Estimation Performs color balancing via
histogram normalization. 1. Determine the histogram for
each RGB channel and find the quantiles that correspond
to our desired saturation level. 2. Cut off the outlying
values by saturating a certain percentage of the pixels to
black and white. 3. The saturated histogram is then scaled
for full range of 0-255.
Results:
A)Input:
Input low light image taken in 800x640 resolution
B) Zero-DCELOW:
Zer-DCELow
light image converted to slight exposure version

C) Zero-DCELargeL:
D) Zero-DCELargeLH:
Performs color balancing via histogram normalization
E)Output Enhanced version:
5. CONCLUSION
We proposed a deep network for low-light image
enhancement. It can be trained end-to-end with zero
reference images. This is achieved by formulating the low-
light image enhancement task as an image-specific curve
estimation problem, and devising a set of differentiable
non reference losses. Experiments demonstrate the
superiority of our method against existing light
enhancement methods. In future work, we will try to
introduce semantic information to solve hard cases and
consider the effects of noise[3]
REFERENCES
[1] Lin Li, Ronggang Wang,Wenmin Wang, Wen Gao, “A
LOW-LIGHT IMAGE ENHANCEMENT METHOD FOR
BOTH DENOISING AND CONTRAST ENLARGING”,
2015,
[2] Li Tao, Chuang Zhu, Jiawen, Tao Lu,Huizhu Jia, “LOW-
LIGHT IMAGE ENHANCEMENT USING CNN AND
BRIGHT CHANNEL PRIOR”, 2017 , IEEE
[3] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy,
Junhui Hou ,” Zero- Reference Deep Curve Estimation
for Low-Light Image Enhancement" ,2020,IEEE
[4] K. Elissa, “Title of paper if known,” unpublished.
Jaemoon Lim, Minhyeok Heo, Chul Lee, Chang-Su Kim,
“Contrast Enhancement of Noisy Low-Light Images
Based on Structure-TextureNoise Decomposition”
,2017 , Journal Visual Communication and
Image Representation
[5] Shi Yangming, Wu Xiaopo,” Low-light Image
Enhancement Algorithm Based on Retinex and
Generative Adversarial Network” , 2019, IEEE
[6] Yu Zhang, Xiaoguang Di, Bin Zhang, Qingyan Li, Shiyu
Yan, “Self-supervised Low Light Image Enhancement
and Denoising” ,2021, Article on arxiv
[7] Anjali Ratate, Priyanka Patil, Siddhi Wavre, “Low
Light Image Enhancement using Convolutional Neural
Network”, 2020, International Research Journal of
Engineering and Technology (IRJET)
[8] Eunjae Ha, Heunseung Lim, Soohwan Yu, “Low-light
Image Enhancement Using Dual Convolutional Neural
Networks for Vehicular Imaging Systems”, 2020, IEEE
International conference on Consumer Electronics
[9] Junyi Xiea, , Hao Biana,, Yuanhang Wua , Yu Zhaoa ,
Linmin Shana , Shijie Haoa, “Semantically-guided low-
light image enhancement” ,2020, International
Association for Pattern Recognition
[10] Linshu Hu , Mengjiao Qin , Feng Zhang , Zhenhong Du
and Renyi Liu, “A CNN-Based Method to Enhance Low-
Light Remote-Sensing Images”, 2021,
Multidisciplinary Digital Publishing Institute
[11] ]https://uwaterloo.ca/vision-image-processing-
lab/research-demos/vip-lowlight-datase

Low Light Image Enhancement Using Zero-DCE algorithm

More Related Content

Similar to Low Light Image Enhancement Using Zero-DCE algorithm (20)

More from IRJET Journal (20)

Recently uploaded (20)

Low Light Image Enhancement Using Zero-DCE algorithm