ISSN (Online) 2278-1021

ISSN (Print) 2319 5940

International Journal of Advanced Research in Computer and Communication Engineering

Vol. 4, Issue 10, October 2015

A Hybrid Processing Approach for Enhancement

of MAV Aerial Images
R Jayadevan 1, Akhila H Das2, Vani M E 3, Athira N P 4, Sreerag K M 5, Jithinraj M V 6
Assistant Professor, ECE Department, Sreepathy Institute of Management & Technology, Vavanoor, India 1
Students, ECE Department, Sreepathy Institute of Management & Technology, Vavanoor, India 2,3,4,5,6

Abstract: This paper proposes a hybrid approach for the enhancement of aerial images obtained from UAV/MAV
cameras. Aerial images are prone to many undesirable effects such as non-uniform lighting, turbidity due to haze and
other atmospheric noise. The aim of the proposed scheme is to achieve the enhancement, where the existing techniques
relying on Wavelet based Dynamic Range Compression (WDRC) algorithm followed by Local Contrast Enhancement
(LCE) fails to achieve desired results. The processing involves both spatial and frequency domain operations which
have been carried out through the implementation of post-processing techniques to those images which have still very
poor contrast after the application of WDRC and LCE. The post processing techniques involve the use of histogram
equalization, removal of noisy artifacts and Laplacian sharpening in a sequential order. The scheme was used for the
enhancement of a series of aerial images obtained from MAV camera on flight. Images enhanced by the scheme are
found to be of much better clarity and vividness, especially in the case of images acquired by MAV inbuilt cameras.
Keywords: Aerial image processing, Laplacian Sharpening, Local contrast enhancement Spatial/Frequency domains,
Wavelet-based Dynamic Range Compression.

I. INTRODUCTION
Unmanned Air Vehicles like UAVS and MAVs have thresholding and filtering. A technique based on digital
small inbuilt cameras to take photographs (aerial images) photogrammetric technique for large scaling mapping has
for a number of civil and defence applications. Images been developed by Udin et al [4], which has advantages
taken during the flight from heights above 100m are like low cost, faster and simplicity. A method for
subjected to various undesirable effects due to camera automatic generation of orthophoto mosaics using scale
movement, non-uniform lighting conditions, turbidity, invariant feature transform (SIFT) for automatic key point
haze, fog, clouds etc. This not only causes a significant detection and matching problems has been proposed by
reduction in the clarity of such images, but the important Marcus et.al in [5].
details may be hidden under the noise that contaminates
Many advanced methods have been developed to improve
the image. It is thus obvious that the utility of these
the local contrast of the images. E. Land’s theory [6] has
captured images, unless processed, is very limited.
produced good results in dynamic range compression and
Natural scenes have high dynamic range. Images captured color constancy while MSRCR (Multi-Scale Retinex with
by the cameras are often poor rendition of the natural Color Restoration) [7] is a retinex based algorithm that
scenes. Even though the human eyes have a lesser uses logarithmic compression and spatial convolution. The
dynamic range, the biological system adjusts to it by main drawback of MSRCR is that the color restoration
dynamic range compression and adapting locally to each changes chromatics of the image in an undesirable
part of the scene. The low dynamic range of cameras manner. Adaptive and Integrated Neighborhood
results in the loss of contrast. There are many algorithms Dependent Approach for Nonlinear Enhancement
for dealing with problems regarding limited dynamic (AINDANE) [8] is used for images taken under low
range like global histogram modification, logarithmic illuminance condition, which includes adaptive luminance
compression etc [1]. However, these methods alone enhancement, adaptive contrast enhancement and color
contribute little to enhancement of the quality of images restoration. Locally Tuned Sine Non-Linear (LTSN)
suffered from the above problems. technique [9] is used for extremely high contrast images,
which includes adaptive intensity enhancement, contrast
Considerable research work is under progress, on the enhancement and color restoration. A novel approach for
processing of images obtained from Micro Air Vehicle removing illuminance effect on aerial images taken under
(MAV), in vehicle detection, remote sensing, and non-uniform lightning condition is proposed Ansari and
correction of camera rolling. Manera et.al [2] presents a Arigela [10], where histogram adjustment and color
computational system which automatically processes restoration are used along with WDRC algorithm [1] to
optical and multispectral MAV images. The system improve the visual quality. But this algorithm (Automatic
includes image acquisition, rectification and image WDRC) is not sufficient for images having very poor
mosaicing. Bharati et.al [3] proposes a differential contrast. In the present work, we propose a post-
morphology closing profile for extracting vehicles processing step which improves the visual quality of those
automatically from traffic images, which includes image images with very poor contrast and eliminates noise from
pre-processing, differential morphology profile, the images.

Copyright to IJARCCE DOI 10.17148/IJARCCE.2015.41035 175

 ISSN (Online) 2278-1021
ISSN (Print) 2319 5940

International Journal of Advanced Research in Computer and Communication Engineering

Vol. 4, Issue 10, October 2015

The algorithm employs a hybrid approach, as it uses both ( ) * ( )+ , - (1)

spatial and frequency domain operations. Initially, the
where ( ) is the intensity of the image at ( )th pixel
histogram of the image is adjusted to avoid the location. The intensity image is then decomposed using
illuminance effect so as to deal with the strong spectral orthonormal wavelet transform as shown in equation (2).
characteristics of the aerial images. Local contrast
enhancement is used in the algorithm in order to restore ( ) ∑ ( ) ∑ ∑ ( )
and improve the contrast lost during dynamic range
compression. A non-linear color restoration is then used to
∑ ∑ ( )
provide color constancy. In the post processing step, use
of histogram equalization enhances the contrast of very ∑ ∑ ( ) (2)
low contrast images for which the approach of Ansari and
Arigela [10] is inadequate. It may also be noted that the
application of Gaussian and median filtering eliminates the
noise in the images which occur while performing color
restoration. Finally, Laplacian sharpening is employed to
make the edges more clearly visible.

II. METHODOLOGY
The proposed algorithm comprises mainly of the following
modules.
(A) Pre-processing
(B) WDRC and LCE
(i) Wavelet based dynamic range compression
(ii) Local contrast enhancement
(iii) Detail coefficient modification
(C) Color Restoration
(D) Post processing
(i) Histogram Equalization
(ii) Removal of noise effect
(iii) Laplacian Sharpening
These modules are discussed in detail in the following
subsections.
A. Pre-processing
Histogram adjustment [10] is performed in the spatial
domain to decrease the illuminance effect occurring due to
non-uniform lightning conditions. This effect causes a
linear shift (after assuming nonlinear properties of human
visual system [11 and 12]) in the beginning of the lower
tail of image histograms and may vary in different spectral
bands. This linear shift can be observed in the histogram
plotted in Figure 1 for a given image corrupted by
atmospheric haze and turbidity along with noise. The shift
due to illuminance effect is determined from the lower
tails of the histogram in Figure 1(c), (e) and (g). For
Channels R, G and B the corresponding shifts are 0.2275,
0.2353 and 0.2039. They are then subtracted from all pixel Figure 1: Histogram adjustment. (a) Input image, (b)
values of the respective channels to form adjusted Histogram adjusted image, (c),(e) & (g) Histograms of
histograms in Figure 1(d), (f) and (h). Figure 1(b) shows RGB channels of input image, (d),(f) & (h) Histograms of
the corresponding histogram adjusted image. RGB channels of histogram adjusted image.
B. WDRC and LCE where are the approximation coefficients at scale
WDRC is basically an algorithm to deal with the problems with corresponding scaling function ( ) and
caused by the limited dynamic range of the imaging are the detail coefficients with corresponding wavelet
devices and to improve its visual quality. It has been function .
modified in which is followed here. A single level decomposition using Haar wavelet is
(i) Wavelet-based dynamic range compression: This employed here for better and low noise performance. This
particular operation is performed on the intensity image in is followed by a scaling of the approximation coefficients
frequency domain. Intensity image is taken as the one with by a two sided raised hyperbolic sine function which pulls
maximum value of the three color channels and is given up the smaller coefficients and pulls down the larger ones
by equation (1) as follows. as performed in equation (3).

Copyright to IJARCCE DOI 10.17148/IJARCCE.2015.41035 176

 ISSN (Online) 2278-1021
ISSN (Print) 2319 5940

International Journal of Advanced Research in Computer and Communication Engineering

Vol. 4, Issue 10, October 2015

( ) the corresponding modified coefficient in ̅ , otherwise the

̅ 0 1 (3)
inverse power as indicated by equation (9). Hence we can
Where eliminate halo artifacts or saturation caused by over
enhancement.
(4) (iii) Detail coefficient modification: The approximation
Equation (4) gives the normalized coefficients and r is the coefficient modification results in loss of details and edge
curvature parameter which adjusts the shape of the deterioration. Thus the image quality is restored by
hyperbolic sine function. Since the approximation appropriate scaling of detail coefficients as shown in
coefficients indicate the dynamic range of the intensity, equation (). So the coefficients are enhanced using
the scaling operation calls for the compression of dynamic equation (11).
range which in turn reduces the image contrast drastically.
Hence it is required to employ a local contrast
enhancement in order to preserve the contrast of the (11)
image. where, and are modified horizontal,
vertical and diagonal coefficients, while and are
(ii) Local contrast enhancement: The local contrast the original horizontal, vertical and diagonal coefficients.
enhancement is done based on centre surround approach. Now the inverse wavelet transform is performed on the
Gaussian filter is used to extract the surrounding intensity modified approximation and detail coefficients to obtain
information. As the first step, normalized approximation the enhanced intensity image .
coefficients are filtered using Gaussian kernel given by
equation (5) C. Color restoration
In order to restore the color of enhanced image, a non-
( ) . / (5) linear approach is employed. The RGB values of enhanced
color image are obtained from the RGB values of original
where is the surround space constant and is obtained color image and the enhanced intensity image as shown in
under the constrain equation (12).
∑ ∑ ( ) (6) ( ) ( ) (12)
An improved rendition is achieved by using a linear ( )
combination of three kernels with three different space where . / , - and is the
( ( )
constants. The values taken for are 4, 40 and80. The nonlinear gain factor, which adjusts the color saturation to
combined scale Gaussian ( ) applied here takes the have a better color rendition.
form.
The resultant image thus obtained after color
( ) ∑ . / (7) restoration is shown in Figure 2(c).

where

Local average image representing the surround is obtained

by 2-D convolution of equation (7) and the elements of
which are (in equation (4)) as given by following
equation.
( )
∑ ∑ ( ) ( ) (8)
where and are the sizes of the Gaussian kernel .
Now the contrast enhanced coefficients matrix is
calculated by equation (9)
Figure 2: WDRC, LCE and Color restoration. (a) Intensity
̅
{ (9) image I of the input image in Figure 1 (a) (b) Enhanced
̅ intensity image I_enh (c) Color restored image I_CR, (d)
where the elements of ̅ are ̅ and is the centre Histogram of color restored image.
surround ratio given by equation (10). The above results are obtained with values for and as
1.2 and 0.2 respectively. The intensity histogram of
( ) (10) reveals that the color restored image has a poor contrast as
the pixel intensity values are concentrated at the middle
where is the enhancement strength constant. is matrix portion of the dynamic range. Moreover, close inspection
which determines whether the centre coefficients are of the Figure 2 (c) indicates a noisy effect resulting from
higher than average surrounding intensity or not. If any color restoration.
value of is less than 1, it is used as a power transform to

Copyright to IJARCCE DOI 10.17148/IJARCCE.2015.41035 177

 ISSN (Online) 2278-1021
ISSN (Print) 2319 5940

International Journal of Advanced Research in Computer and Communication Engineering

Vol. 4, Issue 10, October 2015

In this context, it is highly recommended to adopt a post 0.3103 0.3793 0.3103

processing step to solve the above mentioned issues. = 0.3793 -2.7586 0.3793 (15)
D. Post Processing 0.3103 0.3793 0.3103
From the analysis of intensity histogram in Figure 2(d) and The Laplacian sharpening is implemented according to
the visual inspection of color restored image in Figure equation (16) where ( ) is the final output image.
2(c), some post processing steps are proposed here.
( ) ( ) ( ) (16)
They are Histogram equalization, Laplacian sharpening,
Gaussian filtering and Median filtering. Post processing
methods restore the edges and sharpen the image, since the
image is smoothened during Local contrast
enhancegement. But all the methods are not
simultaneously applicable. In the algorithm, Laplacian
sharpening, Gaussian filtering, median filtering is used for
all images, but histogram equalization is an optional one.
Median filtering removes noise and histogram
equalizations improves the contrast of the image.
(i) Histogram Equalization: As can be seen from Figure
2(d) that the contrast of has to be considerably
increased. This can be achieved easily through a
straightforward approach called histogram equalization. It
is basically a transformation which uniformly distributes Figure 3: Post processing steps (a) Input image (b) Color
the pixel intensity values in the range [0, 1]. restored image I_CR (c) Histogram equalized image, (d)
(ii) Removal of noise effect: Inspection of Figure 2(c) Final output image.
reveals the presence of some common noises like Figure 3 shows the results of post processing steps. Figure
Gaussian and salt and pepper .The post processing step is 3(a) and 3(b) are the input and color restored images
also aimed at removing these noisy artifacts added during respectively. Figure 3(c) shows the histogram equalized
the color restoration step. The filters employed for this can be seen image in which the contrast has been
purpose are Gaussian and Median [13] in a sequential significantly improved, whereas 3(d) represents the final
manner. The application of median filter after Gaussian output image in which the the noisy effects are
filter helps in removing the unwanted impulses in the form considerably reduced along with sharpening the edges.
of salt and pepper if any present in the image. The
Gaussian kernel used for filtering operation is given in III. RESULTS & DISCUSSION
equation (13). The hybrid processing approach has been applied to
several MAV aerial images which suffer from various
( ) (13) undesirable atmospheric effects like non-uniform lighting
In this work, the value of standard deviation used is 0.5 conditions, turbidity, haze, fog, clouds etc. The results thus
and the kernel size used for both Gaussian and median obtained are compared with existing aerial image
filtering is 3x3. Median filtering can preserve enhancement methods like MSRCR, LTSN, AINDANE
discontinuities while it smoothes a few pixels whose and Automatic WDRC. This section also presents a
values differ significantly from their surroundings without discussion on the selection of parameters (curvature
affecting the other. parameter), (enhancement strength constant) and
(nonlinear gain factor).
(iii) Laplacian Sharpening: The application of LCE in step
(B) and noisy removal in step (D) using Gaussian kernels The optimum values of , and obtained empirically
causes some blurring of significant edges and are 1.15, 0.1 and 2 respectively. This selection has been
done through visual inspection of final output images after
discontinuities present in the images. This may affect the
the application of proposed algorithm to around 100 aerial
quality of the final result. This problem can be solved by
images suffering from undesirable atmospheric effects
sharpening the image at the end of the post processing. In
with different combinations of , and . The optimum
this paper, image sharpening is achieved through a well value of is determined only after and have been
known procedure called Laplacian sharpening [13]. optimized. Here we have considered 200 different
Laplacian is a second derivative operator which highlights
combinations of and to optimize their values. Figure 3
gray level discontinuities while it de-emphasizes regionsshows the final output for only four different combinations
with slowly varying intensity levels. The Laplacian of the
of , and . For and , which are their
image obtained after noisy removal ( ) is given by optimum values, 50 different cases of have been taken
equation (14). into account while finalizing its optimum value which is 2.
( ) ( ) It is very obvious from Figure 3 that combination
( ) (14)
corresponding to (d) is visually superior to all the other
The laplacian mask used in our work is given in equation combinations in terms of contrast enhancement, luminance
(15). distribution, color rendition and detail preservation.

Copyright to IJARCCE DOI 10.17148/IJARCCE.2015.41035 178

 ISSN (Online) 2278-1021
ISSN (Print) 2319 5940

International Journal of Advanced Research in Computer and Communication Engineering

Vol. 4, Issue 10, October 2015

Figure 5 shows the performance comparison of proposed

technique with existing aerial image enhancement
methods like AINDANE, LTSN and MSRCR through
visual inspection [14]. The comparison has been done
using 100 MAV aerial images taken at an altitude of
around 70 to 90 meters. Most of the captured images were
subjected to undesirable atmospheric effects like fog,
snow, haze and some noisy artifacts. The captured images
were of size 640x480 in RGB format with a bit depth of
24. Figure 5(a) is an input image with poor contrast and
contaminated with atmospheric noise. The results of
AINDANE, Figure 5 (c), and MSRCR, 5 (e), do not show
a noticeable enhancement. Although LTSN algorithm,
Figure 5 (d), performs relatively well, it lacks a proper
contrast enhancement and color rendition. The result
Figure 4: Optimizing r, α and ρ (a) r=2.5, α=0.4 and ρ=0.5 shown in Figure 5 (f) corresponds to the proposed method
(b) ) r=2, α=0.3 and ρ=3.5 (c) ) r=1.5, α=0.25 and ρ=1 (d) which achieves a significant improvement over other
) r=1.15, α=0.1 and ρ=2 methods. Figure 5(g), (h) and (i), the outputs of
AINDANE, LTSN and MSRCR, do not reveals anything
more than the input image in Figure 5(b). But from the
result in Figure 5 (j), it is very obvious that the proposed
algorithm outperforms the existing ones in terms of
contrast enhancement, luminance distribution, color
rendition, edge sharpening and detail preservation. The
algorithm is implemented in MATLAB 2013a.

IV. CONCLUSION
A new algorithm, incorporating both spatial and frequency
domain analysis for aerial image enhancement is presented
in this paper. The scheme, based on Wavelet-based
Dynamic Range Compression, has been found to be
efficient in enhancing images, particularly aerial images
taken by cameras of MAVs, to desirable levels. The
enhancement to a satisfactory level could be achieved only
through a post processing step which involves histogram
equalization, removal of noise artifacts and Laplacian
sharpening. The enhanced images are found to preserve all
the original details. The algorithm can be easily
implemented and can be used very effectively in a number
of applications of UAVs and MAVs.

ACKNOWLEDGMENT
The authors would like to express their gratitude to
Dr.S.P.Subramanian, Principal and Prof.M.Suhma,
Head, Dept. of Electronics and Communication, Sreepathy
Institute of Management and Technology, Vavannoor,
Kerala for their technical support and constant
encouragement. They are also grateful to Dr. K. T.
Madhavan, Head, R&D, Sreepathy Institute of
Management and Technology for his valuable ideas and
guidance.

REFERENCES
Figure 5: Performance comparison of proposed method [1] N.Unaldi, K.V.Asari, Z.Rahman,”Fast and robust wavelet-based
dynamic range compression and contrast enhancement model with
with other existing aerial image enhancement methods. (a) color restoration..”SPIEproc”2009).
Input image 1.(b) Input image 2. (c), (d), (e) & (f) are the
[2] Manera. J.F, Lucas Rodr’iguez, Claudio Delrieux and Ricardo
final output images of AINDANE, LTSN, MSRCR and Coppo, “Aerial Image Acquistion and Processing for Remote
proposed methods for input image 1. (g), (h), (i) & (j) are Sensing”, JCS&T vol. 10 No. 2.
the final output images of AINDANE, LTSN, MSRCR and [3] Bharti Sharma, Vinod Kumar Katiyar, Arvind Kumar Gupta and
proposed methods for input image 2. Akansha Singh,”The Automated Vehicle Detection of Highway

Copyright to IJARCCE DOI 10.17148/IJARCCE.2015.41035 179

 ISSN (Online) 2278-1021
ISSN (Print) 2319 5940

International Journal of Advanced Research in Computer and Communication Engineering

Vol. 4, Issue 10, October 2015

Traffic Images by Differential Morphological Profile”,Journal of

Transportation Technologies, 2014, 4, 150-156.
[4] Wani Sofia Udin and Anuar Ahmad,”Large Scale Mapping Using
Digital Aerial Imagery of Unmanned aerial vehicle”, International
Journal of Scientific & Engineering Research, Volume 3, Issue 11,
November-2012, ISSN 2229-5518.
[5] Marcus Arthur, Barbados, Raid AL-TAHIR, and Dexter
DAVIS,”Rapid Processing of Unmanned Aerial Vehicles Imagery
for Disaster Management”.
[6] E.Land,”Recent Advances in Retinex Theory.”. Vision
Res.,Vol16,1976,pp 445-458.
[7] ElectronicImaging, January 2004Z. Rahman, D.Jobson and G A
woodwell”, Retinex proessing for Automatic Image Enhancement”.
[8] Tao and K. V.Ansari “An adaptive integrated neighborhood
dependent approach fornonlinear enhancement of color images”,
SPIE Journal of Electronic imaging.
[9] N.Unaldi, S.Temel,Vijayan K Asari and Zia-ur Rahman.,”An
automatic Wavelet-based Nonlinear Image Enhancement Technique
for Aerial imagery”.
[10] K V Ansari, S.Arigela, “An adaptive and nonlinear technique for
enhancement of extremely high contrast images”, IEEE, 2006.
[11] HHulbert, ”Formal connections between lightnessalgorithms”
Jounal of the optical society of America, Vol.3,No 10 pp.1684-
1693(1986).
[12] B. K. P. Horn, “Determining Lightness from an image”, Computer
Graphics and Image Processing, vol. 3, no. 1, pp.277-299, (1974).
[13] R.C Gonzalez,R.E Woods. “Digital image processing using
MATLAB”, 2nd edition Prentice Hall, pp.380-400(2002)
[14] D. J. Jobson, Z. Rahman, G.A. Woodell, “Statistics of visual
representation”, SPIE Proc., Vol. 4736, pp. 25-35. (2002).

BIOGRAPHIES

R. Jayadevan is currently working as

Assistant Professor in Electronics and
Communication Engineering
Department of Sreepathy Institute of
Management and Technology,
Vavanoor, Palakkad. He is doing part-
time Ph. D in Government Engineering
College Thrissur in the area of image
processing. He received his M.Tech degree from
Government Engineering College, Thrissur, University of
Calicut in 2012 and B. Tech degree from Jyothi
Engineering College, Cheruthuruthy, University of Calicut
in 2010. He has published seven papers including
international journal and conferences. He has also
published two books. His area of specialization includes
signal processing and image processing.

Authors Akhila H Das, Vani M E, Athira N P, Sreerag

K M , Jithinraj M V received their B. Tech degree from
Sreepathy Institute of Management and Technology,
Vavanoor, University of Calicut. Their area of interest is
image processing.

Copyright to IJARCCE DOI 10.17148/IJARCCE.2015.41035 180