Image processing
using fusion based
enhancing method
Presented by
Supervisor J.Jayasree(960217106058),
Dr. R.Ablin P.Mary Monica(960217106085),
Associate Professor/ECE
M.Nithya Sri(960217106097).
OVERVIEW
• Objective
• Abstract
• Introduction
• Literature review
• Proposed method
• Experimental Results
• Conclusion
• Future Scope
• References
OBJECTIVE
To enhance a weakly illuminated images with weakly illuminated
conditions such as backlighting, non-uniform illumination and
nighttime using image fusion method.
ABSTRACT
• An illumination estimating algorithm is used to decompose an
observed image into a reflectance image and an illumination image.
• Based on these images, two inputs were derived using the sigmoid
function and adaptive histogram equalization.
• Based on these inputs two weights were designed which will produce
an adjusted illumination.
• Final Enhanced image is obtained by compensating the adjusted
illumination back to the reflectance.
INTRODUCTION
• In weakly illuminated environments, the quality of images and video captured
by optical imaging devices are often degraded.
• In this process, a fusion-based method is used for enhancing weakly illuminated
images with weakly illuminated conditions such as backlighting, non-uniform
illumination and nighttime.
• First, our method is applicable to singles image which differ from several fusion
techniques that require multiple images of the same scene to perform fusion.
• Second, a simple illumination estimating algorithm based on morphological
closing is performed which estimates naturalness and luminance.
• Third, an effective fusion-based algorithm is performed to adjust the estimated
illumination.
LITERATURE REVIEW
S. No. Method(Year) Concept Advantages Disadvantages
1 Variational Performed by Preserved Spatial Large gray level
Method(2017) histogram mapping Information differences with very dark
regions.
2 Generalized Unsharp Decompose an Noise removal and edge Weak color information is
masking image into high- enhancement often a problem.
Method(2017) frequency and low-
frequency
3 Naturalness- Design a band pass Illumination can be Non-uniformly
preserving filter to estimate adjusted illuminated images
enhancement illumination and
Method(2018) propose a bi-log
transformation
4 Variational Contrast Uses Spatial Suppresses over Unnatural images and
Enhancement information to enhancement artifacts.
Method(2018) preserve the shape of
the input histogram
4 Histogram based Generates an output Enhances low contrast Contrast Over
Method(2018) with a desired Shape images enhancement and noise
amplification
5 Filtering based Compresses high Not intended for Computationally
Method(2019) dynamic range using backlighting and nighttime intensive
local linear filtering illuminated images.
6 Retinal based Image is taken to be a Estimates an enhanced Does not apply on multi
Method(2019) product of reflectance result, Used for both color resolution images.
and illumination. and grey scale images.
Reflectance is
considered to be the
enhanced image.
PROPOSED METHOD
Derived inputs Inputs - Laplacian pyramid Weights - Gaussian
pyramid
*
Illumination
* Å
Original Image
Multi-scale fusion
Reconstruct
*
Reflectance Adjusted illumination
Enhanced Image
Block diagram of image fusion method
• Our goal is to develop a fast and effective fusion method that blends the
advantages of different mature image processing techniques.
• First, we propose an algorithm for estimating illumination based on a
morphologically closing operation, the filtering operation requires a high
computational time, which is not suitable for practical applications.
• According to the retinex theory, a captured image is the product of reflectance
and illumination written as,
=I ……………………..(1)
Where S is the measured image,
R is the reflectance,
I is the illumination,
c is the colour channel of RGB (red, green, blue) and
(x,y) is the pixel location
• Many techniques exist for estimating the illumination and reflectance layer
Inspired by the dark channel prior ,which calculates the local minimum in the
three color channels for image dehazing ,here proposed the following
illumination estimating algorithm .
• First we obtain an image lightness from the maximum value of its three
color channels to represent luminance variance, which we can write as
L= ……………………..(2)
• This type operation is the reverse of the first step for obtaining the dark
channel ,which finds the minimum of the three color channels
• Second, since the illumination is locally smooth, some filter operations
should be adopted to refine L.
• To achieve both local smoothness and computational efficiency, a simple but
effective algorithm for illumination estimation by using a morphologically
closing operation:
I =L∙P/225 ……………….........(3)
Where P denotes the structuring element and ∙ is the closing operation.
Diving by 225 is used to map the range to [0,1] for the downstream
operations.
• The illumination can be effectively estimated by the equation (3).
• It refines the estimated illumination with the guided filter to preserve the
shape of large contours.
• The filtering output is expressed as a weighted average at a pixel i:
......................(4)
Where i and j are pixel indexes, W is the filter kernel that is a function of
the guidance image g.
• The guidance image uses the kernel
…….................(5)
window, 𝞵𝑘 and are the mean and variance of g, 𝛆 is a regularization parameter .
Where is a window centred at the pixel k andis the number of pixels in the
• In this paper, the guidance image g is the V layer in HSV color space of the input
image S.
• In our proposed fusion-based approach ,three inputs are derived from the estimated
illumination I.
• The first input ,is the original estimated illumination I
• The second input is designed to address the global luminance in order to help clarify
the dark regions of the image .
• Many enhancing operators and functions can be adopted to improve global
luminance ,such as gamma correction and the sigmoid function .
• The second input by using the arc tangent transformation of I:
……………………(6)
Here 𝜆 is a parameter that controls the degree of luminance .
• Multiplication by 2/π maps the range to [0,1).
• Thus an automatic 𝜆 setting is introduced based on the
experiments: …….……………….(7)
• It is obvious that a smaller Imean indicated darker luminance ,and thus a higher 𝜆 is
obtained to make a higher level of enhancement
• Designing weights for fusing the three input images .
• The standard deviation and the histogram of the illumination values for each of
2000 images has to be estimated.
• The mean approximately equals 0.5 and the mean of the standard deviations,
approximately equals 0.5 and the mean of the standard deviations, approximately
equals 0.25 after mapping the range to[0,1]
• Based on the observations, set the brightness weight to be,
…………………………(8)
• Where k denotes the kth derived input.
• A second weight, which we call the chromatic contrast weight ,evaluates the
contrast by combining the estimated
• Illumination with the chromatic information .
• This weight is computed based on the chromatic filtering formula
=(x,y)(1+cos(αH+ф)S ) ……………………..( 9)
• Where H is the hue and S is the saturation in HSV color space of the original
input color image .
• To calculate the third weight
= …..……………………(10)
• Then normalize this to give the final weight,
..........................(11)
• Then to obtain the adjusted illuminations as follows,
(x,y(x,y)(x,y) …………….............( 12)
• Since the sum of each equals to 1, the result’s intensity scale is guaranteed to
be the same as the inputs
• Based on the Laplacian operator ,each input image is represented as the sum of
patterns computed at different scales.
• The inputs are convolved by a Gaussian kernel to generate a low pass filtered
versions.
• Decompose each derived input into a Laplacian pyramid to extract image
features, and each normalized weight into a gaussian pyramid to smooth the
strong transitions.
• Both the Gaussian and Laplacian pyramids have the same number of levels.
• Then fuse pyramids by mixing each level of the Gaussian and Laplacian
pyramids ,written as
(x,y ……….…………( 13)
Where l is the number of the pyramids levels, set equal to 6,is the Laplacian
versions of the derived input and is the Gaussian version of the normalized map .
• The refined adjusted illumination is computed by summing each level of the
fused pyramid,
I ……………………(14)
• Where is the up sampling operator with factor d=
• The final enhanced color image is generated by compensating the adjusted
illumination back to the reflectance
= ………………………( 15)
• The results of our fusion approach using this multi-scale technique.
EXPERIMENTAL RESULTS
• In this original image a bright
background and dark fore ground
exists.
• Certain details and objects in dark
region cannot be captured clearly
while the image has a wide
dynamic range.
• After applying our image fusion
based method result of the street
image improves its luminance and
Original Image
the details can be seen clearly.
• Hence our proposed method
enhances the dark fore ground
while maintaining the bright
background without introducing
artifacts.
• The results are also too bright
for visual perception especially
in the night time image.
• The elapsed time for our
method is 22.354466
Enhanced Image
CONCLUSION
• An illumination estimating algorithm is first proposed to extract
luminance and represent naturalness.
• By choosing appropriate inputs and weights from the estimated
illumination, the proposed framework can effectively deal with
images under different illumination conditions such as low light
images, backlight images and non-uniformly illuminated
images .
• Experimental results demonstrate that the proposed algorithm
generates high quality images with improved luminance and
contrast enhancement.
Future Scope
In future by using post processing techniques can improve the
visual qualities and other deficiency in an image.
REFERENCES
[1] C. Lee, C.Lee, C.S.Kim, Contrast enhancement based on layered difference
representation of 2d histograms, IEEE Trans. Image Process. 22(12) (2013)
5372–5384.
[2] R.C. Gonzalez, R.E. Woods, Digital Image Processing, Prentice-Hall,
Englewood Cliffs, NJ, USA, 2007.
[3] Lee. C, Lee. C, Lee. Y. Y, (2012) “Power-constrained contrast enhancement
for emissive displays based on histogram equalization”, IEEE Trans. Image
Process. 21 (1) 80–93.
[4] T.Arici, S.Dikbas, Y.Altunbasak, A histogram modification frame work and
its application for image contrast enhancement, IEEE Trans. Image
Process.18(9) (2009)1921–1935.
[5] Khann. N, Arya. K, Pattanaik. M, (2013) “Histogram statistics-based
variance controlled adaptive threshold in anisotropic diffusion for low contrast
image enhancement”, Signal Process. 93 (6) 1684–1693.
[6] G. Deng, A generalized unsharp masking algorithm, IEEE Trans. Image
Process. 20 (5) (2011) 1249–1261
[7] L.Yuan, J.Sun, Automatic exposure correction of consumer photographs, in:
Proceedings of European Conference on Computer Vision, 2012, pp.771–785.
[8] Wang. L, Xiao. L, Liu. H, (2014) “Variational Bayesian method for
Retinex”, IEEE Trans. Image Process. 3381–3396.
[9] Yuan. L, Sun. J, (2012) “Automatic exposure correction of consumer
photographs, in: Proceedings of European Conference on Computer Vision”,
pp. 771–785.
[10] Zhou. Z, Li. S, Wang. B, (2014) “Multi-scale weighted gradient-based
fusion for multi- focus images”, Inf. Fusion 20, 60–72.
Thank you
