International Journal of Computer Applications (0975 – 8887)

Volume 156 – No 10, December 2016

Weighted Guided Image Filtering – A Survey

Nidhi Sen Akhilesh Jain
Mtech scholar, Asst. Prof.,
Department of Department of Electronics & Communication
Electronics & Communication Engineering Engineering
NRI Institute of information NRI Institute of information
science & technology science & technology
Bhopal, India Bhopal, India

Swapnil Jain
H.O.D, Department of Electronics &
Communication Engineering
NRI Institute of information
science & technology
Bhopal, India

ABSTRACT anisotropic diffusion with a nontrivial stable state. The

It is well-known that local filtering-based edge preserving guidance image is able to be another image besides the
smoothing method suffers from halo artifacts. In this paper, a filtering input in several purposes. For illustration, in
weighted guided image filter is introduced by incorporating colorisation the chrominance channels be supposed to not
an edge-aware weighting into an accessible guided image bleed across luminance edges; in image matting the alpha not
filter to address the problem. The WGIF inherit benefits of shiny should capture the thin structures in a combined image;
both global and local smoothing filters in the sense that: 1) the in haze removal the depth layer should be consistent with the
difficulty of the WGIF is O(N) for an image with N pixels, outlook. In these cases, we look upon the
which is same as the GIF and 2) the WGIF can avoid halo chrominance/alpha/depth layers as the image to be filtered,
artifact like the existing global smoothing filters. The WGIF and the composite /luminance/scene as the guidance image,
is applied for single image detail enhancement, single image correspondingly. The filtering process is attained by
mist removal, and fusion of differently exposed images. optimizing a quadratic rate function weighted by the guidance
Investigational results show that the resultant algorithms image. The result is known by solving a large sparse matrix
create images with better visual quality and at the same time which only depends on the guide. This inhomogeneous matrix
halo artifacts can be avoided from appearing in the final implicitly describes a translation-variant filtering kernel.
images with negligible rise on running times. While these optimization based approach frequently defer
state-of-the-art class, it comes with the price of costly
Keywords computational time. One more way to take benefit of the
Edge-preserving smoothing, weighted guided image filter, guidance image is to openly build it into filter kernels. The
edge-aware weighting, detail enhancement, haze removal, two-sided filter, independently wished-for in, and and later
exposure fusion . indiscriminate in, is perchance the most accepted one of like
explicit filters. Its output at a pixel is a weighted average of
1. INTRODUCTION the in close proximity pixels, where the weights depend on
MOST application in computer vision and computer graphics the color /intensity similarities in the guidance image. The
engages image filtering to restrain and/or remove content in guidance image itself can be the filter input or another image.
images. plain linear translation-invariant filters with explicit The two-sided filter can smooth small fluctuations and while
kernels, like mean, Gaussian, Laplacian, and Sobel filters, preserving edges. However this filter is valuable in various
have been broadly used in image restoration, states, it may have unnecessary gradient reversal artifacts near
sharpening/blurring, feature extraction, edge detection etc. edges. The speedy execution of the bilateral filter is also a
Alternatively, LTI filters be capable of implicitly performed challenging problem. Latest techniques rely on quantization
by solving a Poisson Equation the same as in high dynamic process to accelerate but may sacrifice accuracy. In this paper,
range (HDR) compression, image matting, image stitching, we propose a novel precise image filter called guided filter.
and gradient domain manipulation. The filtering kernels are The filtering output is nearby a linear transform of the
absolutely described by means of the inverse of a guidance image. On one side, the guided filter has fine edge-
homogenous Laplacian matrix. The LTI filtering kernels are preserving smoothing property like the bilateral filter, but it
spatially independent and invariant of image content. But does not experience from the gradient reversal artifacts. On
usually one may desire to consider added information from a the other side, the guided filter is capable of to be used
known guidance image. The prepare work of an isotropic beyond smoothing: With the help of the guidance image, it
distribution uses the incline of the filtering image itself to can create the filtering output more structured and less
direct a diffusion process, steer clear of smoothing edges. The smoothed as compared to the input. We notice that the guided
weighted least squares filter utilizes the filtering input (as an filter performs extremely well in a great range of applications,
alternative of intermediate results, as in) as the guidance, and including image smoothing and enhancement, HDR firmness,
optimizes a quadratic function, which is the same to flash-no-flash imaging, matting, feathering, dehazing, and

29
 International Journal of Computer Applications (0975 – 8887)
Volume 156 – No 10, December 2016

joint up sampling. Moreover, the guided filter naturally has a local image structure, classified in line with the output of the
time (in the number of pixels N) non -estimated algorithm for Laplacian of Gaussian (LoG) operator applied to the degraded
both high resulation images and gray-scale, regardless of the image. The optimal and are estimated in advance throughout
kernel size and intensity range. Classically, our CPU the development of the algorithmic rule via an off-line
execution achieves 40 ms per mega-pixel executing gray- training procedure employing a set of pairs of original and
scale filtering: To the best of our knowledge, this is one of the degraded images. During the training procedure, the MSE
best edge preserving filters. A beginning description of this between the original and reconstructed images is reduced for
paper was published in ECCV ’10. It is significance each class of pixels.
mentioning that the guided filter has witnessed a sequence of
new applications since then. The guided filter allows a high- Anush Krishna Moorthy et. al. [3] ―A Two-Step Framework
quality real-time stereo matching algorithm. An alike stereo for Constructing Blind Image Quality Indices‖ Now a day’s
technique is proposed independently in the guided filter has No reference Image Quality Assessment (NR IQA)
also been applied in optical flow assessment, interactive algorithms usually assume that the distortion affecting the
image segmentation, saliency recognition, and illumination picture is recognized. This is a preventive supposition for
rendering. We accept as true that the guided filter has realistic applications; since in a majority of cases the
enormous potential in computer vision and graphics, given its distortions the picture are unidentified.We recommend a
ease, effectiveness, and high-quality. We have provided a innovative 2-step framework for no-reference image quality
public code to make easy future studies. assessment based on natural scene statistics (NSS). Previously
trained, the frameworks don’t require any knowledge of the
2. LITERATURE SURVEY distorting process and the framework is modular in that it can
Zhengguo Li et. al. [1] ―Weighted Guided Image Filtering‖ be extended to any number of distortions. We describe the
In this paper, It is identified that local filtering-based edge framework for blind image quality consideration and a
preserving smoothing method suffer from halo artifacts. In description of this framework—the blind image quality index
this paper, a weighted guided image filter is initiated by (BIQI) is evaluated on the LIVE image quality assessment
incorporating an edge-aware weighting into an accessible database.
guided image filter (GIF) to deal with the problem. The Jiahao Pang et. al. [4] ―Improved Single Image Dehazing
WGIF inherits benefit of both global and local smoothing
filters in the sense that: 1) the difficulty of the WGIF is O(N) Using Guided Filter‖ Single image dehazing is challenging
for an image with N pixels, which is similar as the GIF and 2) because it is massively ill-posed. Haze removal based on dark
the WGIF can keep away from halo artifacts like the present channel prior is effective, but refining the transmission map
global smoothing filters. The WGIF is applied for single with closed-form matting is computationally expensive.
image detail improvement, single image haze deletion, and Recent work discovered that using guided filter to refine the
blend of differently exposed images. A weighted guided transmission map is feasible. In this paper, we intricate single
image filter is planned in this paper by incorporating an edge- image dehazing by combining dark channel prior and guided
aware weighting into the guided image filter .The WGIF image filtering in detail. By analyzing the tradeoffs of this
preserves keen boundaries as well as presented global filters, approach, we tend to propose an effective scheme to adapt the
and the difficulty of the WGIF is O (N) for an image with N parameters Experiments and comparisons show that our
pixels which is almost same as GIF. Because of the simplicity method generates satisfactory dehazed results with low
of the WGIF, it has several applications in the field of computation. In this paper, we elaborate single image
computational photography and image processing. Mainly, it dehazing by combining dark channel prior and guided image
is useful to study single image aspect enhancement, fusion of filtering, and then we study several aspects of this approach.
differently exposed images, and single image haze removal. Through experiments and analyses, we propose an effective
Investigational results illustrate that the resulting algorithms scheme to adapt the patch size of dark channel and the
can produce images with excellent visual quality as those of filtering radius of guided filter. The main benefit of using
global filters, and same time the running times of the planned guided filter to refine the transmission lies in its low
algorithms are comparable to GIF based algorithms. computational cost; it also generates comparable dehazed
results with He et al.’s work. Since guided image filtering is
Buyue Zhang et. al. [2] ―Adaptive Bilateral Filter for an approximation of soft matting in, this method may fail
Sharpness Enhancement and Noise Removal‖ In this paper, when the input image contains abrupt depth changes.
we tend to present the adaptive bilateral filter (ABF) for Fortunately, it turns out that our work performs quite well on
sharpness improvement and noise removal. The ABF sharpen many hazy images, its O (N) time complexity also makes it
an representation by raise the incline of the edges while not appealing for many applications.
manufacturing overshoot or undershoot. It’s an approach to
sharpness improvement that's basically different from the Pierre Charbonnier et. al. [5] ―Deterministic Edge-Preserving
unsharp mask (USM). This new approach to slope restitution Regularization in Computed Imaging‖ Many image process
also differs significantly from previous slope restoration problems are sick posed and must be regularised. Usually, a
algorithms in this the ABF doesn't involve detection of edges roughness penalty is imposed on the explanation. The
or their direction, or pulling out of edge shape. In the ABF, complexity is to evade the smooth of edges, which are
the edge slope is enhanced by reworking the histogram important attributes of the image. In this paper, we tend to 1st
through a variety of filter with adaptive offset and width. we provide conditions for the design of such an edge-preserving
tend to demonstrate that ABF works well for both natural regularization. Under these conditions, we tend to show that
pictures and text images. In this paper, we tend to there an it's possible to introduce an auxiliary variable whose role is
adaptive bilateral filter (ABF). The ABF retains the general twofold. First, it marks the discontinuities and makes sure
kind of the bilateral filter, but contains 2 necessary their perpetuation from flat. 2nd it makes the criterion half-
modifications. First, an offset is introduced to the range filter quadratic. The optimisation is then easier. We tend to propose
within the ABF. Second, both and also the width of the range a deterministic strategy, based on alternate minimizations on
filter in the ABF are locally adaptive. They depend upon the the image and also the auxiliary variable. This results in the
definition of an ingenious reconstruction algorithmic rule,

30
 International Journal of Computer Applications (0975 – 8887)
Volume 156 – No 10, December 2016

referred to as ARTUR. A number of theoretical property of The O (N) times Edge-Avoiding Wavelets (EAW) is wavelets
ARTUR are discuss. Experimental results illustrate the with precise image-adaptive weights. But the kernel of the
behavior of the algorithmic rule. These results are shown in wavelets are sparsely distributed in the image plane, with
the field of tomography, but this technique can be applied in a constrain kernel range (to powers of two), which may limit
large number of applications in image processing. In this the applications. Recently, Gastal and Oliveira recommend a
document, we've measured the trouble of edge-preserving auxiliary O (N) time filter acknowledged as the ―Domain
Regularization in computed imaging. Our first aim was to Transform filter‖. The key idea is to iteratively and separably
make available a united answer to the question, what apply 1D edge-aware filter. The O (N) time complexity is
belongings have to a possible work (or its derivative) satisfied achieved by integral images or recursive filtering. We
to make sure the maintenance of edges. We’ve planned a determine evaluate with this filter in this paper.
heuristical study of the first-order necessary conditions which
LED us to propose 3 conditions for edge preservation. B. Implicit Weighted-Average Filters
A string of approach optimizes a quadratic cost function and
3. METHOD solves a linear system, which is equivalent to implicitly filter
We are analysis edge-preserving filtering method in this a figure by an inverse matrix. In image segmentation and
section. We categorize them as implicit/explicit weighted- colorization, the affinities of this matrix are Gaussian
average filters and non average ones: functions of the shade similarity. In image matting, a matting
Laplacian matrix is designed to enforce the alpha matte as a
A. Explicit Weighted-Average Filters confined linear transform of the image colors. This matrix is
The two-sided filter is perhaps the simplest and most also applied in haze removal. The weighted least squares filter
instinctive one among precise weighted-average filters. It in adjust the matrix resemblance according to the image
computes the filtering output at each pixel as the average of gradients and produces halo-free edge-preserving smoothing.
adjacent pixels, weighted by the Gaussian of both intensity Although these optimization-based approaches often generate
and spatial distance. The bilateral filter smooths the image high quality results, solving the linear arrangement is
whereas preserving edges. It has been broadly used in noise prolonged. Direct solution like Gaussian Elimination is not
reduction, HDR compression, multi scale detail decay, and practical due to the memory-demanding ―filled in‖
image abstraction. It is generalized to the joint bilateral filter predicament. Iterative solvers like the Jacobi method,
in, where the weights are computed from one more guidance Successive Over-Relaxation (SOR), and Conjugate Gradients
image moderately than the filtering input. The joint bilateral are too slow to converge. Though carefully designed pre-
filter is particularly preferential when the image to be filtered conditioners greatly reduce the iteration number, the
is not dependable to provide edge information, e.g., when it is computational cost is still high. The multi grid method [is
very loud or is an in-between result, such as in flash/no-flash proven O (N) time complex for homogeneous Poisson
de noising, image up sampling, image deconvolution, stereo equations, but its quality degrades when the matrix become
matching, etc. The bilateral filter has limits despite its extra inhomogeneous. Empirically, the implicit weighted-
attractiveness. It has been become aware of in that the average filters take at least a few seconds to progression a
bilateral filter might experience from ―gradient reversal‖ single megapixel picture either by pre-conditioning or by
manufactured article. The cause is that when a pixel has few multigrid. It has been observed that these inherent filters are
similar pixels around it, the Gaussian is the efficiency very much related to the explicit ones. In Elad shows that the
.Abrute-force implementation is O (Nr2) time with kernel bilateral filter is one Jacobi iteration in solution of the
radius r. Durand and Dorsey recommend a piece-wise linear Gaussian affinity matrix. The Hierarchical Local Adaptive
depiction and help FFT-based filtering. Paris and Durand Preconditioners and the Edge-Avoiding Wavelets are
formulate the gray-scale bilateral filter as a 3Dimensional constructed in a analogous mode. In this paper, we explain
filter in a whole variety domain, and down sample this area to that the guided filter is closely related to the matting
speed up if the Nyquist condition is something like true. In the Laplacian matrix.
case of package spatial kernels, Weiss proposes an O (N logr)
time method based on distributive histograms, and Porikli C. Non average Filters
proposes the first O (N) time method using integral Edge-preserving filter can also be attained by non average
histograms. We point out that constructing the histograms is filters. The median filter is a well-known edge-aware
effectively perform a 2D spatial filter in the space-range operator, and is an exceptional case of narrow histogram
domain with a 1D range filter followed. Under this viewpoint, filters. Histogram filters have O (N) time implementations in
both and sample the signal along the range domain but do not a way as the bilateral grid. The Total-Variation (TV) filter
reconstruct it. Yang proposes another O (N) time method optimizes an L1-regularized cost function, and are shown
which interpolates all along the range domain to allocate more equivalent to iterative median filtering. The L1 rate function
aggressive sub sampling. All of the above methods are may also be optimized by means of half-quadratic split,
linearly difficult w.r.t. the numeral of the sample intensities alternating between a quadratic model and spongy contraction
(e.g., number of linear pieces or histogram bins). They require (thresholding). In recent times, Paris et al. proposed
common sampling to attain acceptable speed, but at the manipulating the coefficients of the Laplacian Pyramid in the
expense of quality degradation if the Nyquist condition is region of every pixel for edge-aware filtering. Xu et al.
harshly broken. The space-range domain is indiscriminate to Propose optimizing an L0-fixed cost function favoring
higher dimension for color-weighted bilateral filtering. The piecewise steady solution. The non average filters are often
exclusive cost owing to the elevated dimensionality can be computationally expensive
reduced by the Gaussian kd-trees, the Permutohedral Lattices,
or the Adaptive Manifolds. But the performance of these 4. CONCLUSION
methods is not competitive for grayscale bilateral filters This paper has reviewed the mainly latest research trends and
because they use up a lot additional time preparing the data proposed a weighted guided image filter (WGIF). In this
structures. Given the limitations of the bilateral filter, people paper WGIF is introduced by incorporating an edge-aware
began to explore new designs of fast edge-preserving filters. weighting into the guided image filter (GIF). In this paper we

31
 International Journal of Computer Applications (0975 – 8887)
Volume 156 – No 10, December 2016

have presented the survey of weighted guided image filtering. [8] K. Moorthy and A. C. Bovik, ―A two-step framework for
In this paper proposed that local filtering-based edge constructing blind image quality indices,‖ IEEE Signal
preserving smoothing techniques. Process. Lett., vol. 17, no. 5, pp. 513–516, May 2010.

5. REFERENCES [9] L. I. Rudin, S. Osher, and E. Fatemi, ―Nonlinear total

[1] Zhengguo Li, Jinghong Zheng, Zijian Zhu, Wei Yao, variation based noise removal algorithms,‖ Phys. D,
Shiqian Wu ―Weighted Guided Image Filtering‖ IEEE Nonlinear Phenomena, vol. 60, nos. 1–4, pp. 259–268,
Transactions on Image Processing, Vol. 24, No. 1, Nov. 1992.
January 2015 1057-7149 © 2014 IEEE. [10] Z. G. Li, J. H. Zheng, and S. Rahardja, ―Detail-enhanced
[2] B. Y. Zhang and J. P. Allebach, ―Adaptive bilateral filter exposure fusion,‖ IEEE Trans. Image Process., vol. 21,
for sharpness enhancement and noise removal,‖ IEEE no. 11, pp. 4672–4676, Nov. 2012.
Trans. Image Process., vol. 17, no. 5, pp. 664–678, May [11] R. Fattal, M. Agrawala, and S. Rusinkiewicz,
2008 ―Multiscale shape and detail enhancement from multi-
[3] A. K. Moorthy and A. C. Bovik, ―A two-step framework light image collections,‖ ACM Trans. Graph, vol. 26, no.
for constructing blind image quality indices,‖ IEEE 3, pp. 51:1–51:10, Aug. 2007.
Signal Process. Lett. vol. 17, no. 5, pp. 513–516, May [12] R. C. Gonzalez and R. E. Woods, Digital Image
2010. Processing. Upper Saddle River, NJ, USA: Prentice-Hall,
[4] J. Pang, O. C. Au, and Z. Guo, ―Improved single image 2002.
dehazing using guided filter,‖ in Proc. APSIPA ASC, [13] Levin, D. Lischinski, and Y. Weiss, ―A closed-form
Xi’an, China, 2011, pp. 1–4. solution to natural image matting,‖ IEEE Trans. Pattern
[5] P. Charbonnier, L. Blanc-Feraud, G. Aubert, and M. Anal. Mach. Intell., vol. 30, no. 2, pp. 228–242, Feb.
Barlaud, ―Deterministic edge-preserving regularization 2008.
in computed imaging,‖ IEEE Trans. Image Process., vol. [14] J. Preetham, P. Shirley, and B. Smits, ―A practical
6, no. 2, pp. 298–311, Feb. 1997. analytic model for daylight,‖ in Proc. SIGGRAPH, 1999,
[6] Pierre Charbonnier, Laure Blanc-F´eraud, Gilles Aubert, pp. 91–100.
Michel Barlaud ―Deterministic Edge-Preserving [15] S. G. Narasimhan and S. K. Nayar, ―Chromatic
Regularization in Computed Imaging‖ IEEE framework for vision in bad weather,‖ in Proc. IEEE
TRANSACTIONS ON IMAGE PROCESSING, VOL. Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun.
6, NO. 2, FEBRUARY 1997 2000, pp. 598–605.
[7] T. Mertens, J. Kautz, and F. Van Reeth, ―Exposure
fusion: A simple and practical alternative to high
dynamic range photography,‖ Comput. Graph. Forum,
vol. 28, no. 1, pp. 161–171, 2009.

IJCATM : www.ijcaonline.org
32