Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

IETE Journal of Research

ISSN: 0377-2063 (Print) 0974-780X (Online) Journal homepage: https://www.tandfonline.com/loi/tijr20

Multifocus Image Fusion Based on Multiresolution


Pyramid and Bilateral Filter

Simrandeep Singh, Nitin Mittal & Harbinder Singh

To cite this article: Simrandeep Singh, Nitin Mittal & Harbinder Singh (2020): Multifocus Image
Fusion Based on Multiresolution Pyramid and Bilateral Filter, IETE Journal of Research, DOI:
10.1080/03772063.2019.1711205

To link to this article: https://doi.org/10.1080/03772063.2019.1711205

Published online: 15 Jan 2020.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tijr20
IETE JOURNAL OF RESEARCH
https://doi.org/10.1080/03772063.2019.1711205

REVIEW ARTICLE

Multifocus Image Fusion Based on Multiresolution Pyramid and Bilateral Filter

Simrandeep Singh1 , Nitin Mittal1 and Harbinder Singh2


1 Department of Electronics & Communication Engineering, Chandigarh University, Gharuan, India; 2 Department of Electronics &
Communication Engineering, Chandigarh Engineering College, Landran, India

ABSTRACT KEYWORDS
Multifocus image fusion is the process for enhancing the perception of a vision by combining sub- Bilateral filter; Detail layer;
stantial information captured by the camera from different focus points. It is very difficult to obtain Gaussian pyramid; Laplacian
a finer focus for all objects in the scene. The solution to the problem is to capture several multifocus pyramid; Multifocus images
images with different focal points, which are merged to obtain a single fused image with improved
Depth of Field (DOF). The proposed method utilizes the unique significant texture features that are
obtained from the Bilateral Filter (BLT). These unique features are used to compute weight map
function that controls the transformation of details from multifocus source images to a single fused
image. A pyramid-based multiresolution decomposition is proposed for analyzing spatial frequency
components and then the fused image is reconstructed from modified spatially band-passed images
by applying an inverse multiresolution transform. The fused image yields better results in terms of
entropy, mean, standard deviation, average gradient, and PSNR values over existing state-of-the-art
multifocus image fusion techniques. The proposed algorithm retains the actual shape of objects and
also avoids the introduction of visible artifacts in the fused image.

1. INTRODUCTION are very clear, more detectable, and more recognizable.


Nowadays, digital image processing is used in various Multifocus image fusion is a process of joining multi-
application domains. Because of the inherent drawbacks ple images acquired from the same camera focused at
associated with the digital camera and its image quality, different objects to enhance the various properties of an
there has been a great scope for developing techniques image [4]. In various machine vision, medical imaging,
to enhance the picture quality. Image fusion is a proce- and remote-sensing applications, it is desirable to over-
dure where several images can be pooled together to form come the limitation of ray optics used in the camera,
a single fused image [1]. These several source images which focuses on different objects while clicking from
can be acquired by different imaging sensors at different different distances. The most significant matter concern-
wavelengths while concurrently viewing the same sight ing multifocus image fusion is to decide how to select
[2]. The resultant combined image is made to improve the in-focus part across source images and how to com-
image content. People can detect, recognize, and iden- bine them to produce a single output image. In practice,
tify targets in the fused image easily [3]. Moreover, the all fusion algorithms perform the image fusion process
fused image contains more accurate information about directly on the source multifocus images. A very sim-
the scene, which may be more suitable for computer pro- ple spatial domain image fusion method includes where
cessing. Images captured using the same camera may also gray level average is taken pixel-by-pixel from the source
vary in numerous ways and a single image out of them is images [6,7]. The biggest drawback of this approach is
not enough to analyze the exact situation [4]. The images the reduction of contrast [8]. After the evolution of pyra-
captured using the same sensor may vary in time coor- mid transform, some more complex approaches emerge
dinates, so to combine multiple images, it is necessary to out [9]. If fusion is carried out in the transform domain,
bring the input images to a common coordinate system. it produces better results. The most useful technique is
The process of geometrically aligning two or more images the pyramid transform for this purpose. In pyramid, a
is called image registration [5]. It is not the scope of this representation image is expressed as a summation of spa-
research paper, but it needs to be discussed here because tially band-passed images where each band retains local
an assumption is used here that either all source images spatial information [10]. On each source image, multires-
are already registered or acquired with the help of tripod olution decomposition is performed, and then all these
from the static scene. The composite image has a large decomposed bands are integrated to form a composite
DOF (Depth of Field) and contains all the objects that representation. The final reconstructed fused image is
© 2020 IETE
2 SIMRANDEEP SINGH ET AL.: MULTI-FOCUS IMAGE FUSION

obtained by performing an inverse multiresolution trans- separate fine details from source images, which are fused
form [11]. While going through the literature of pyramid separately. Moreover, these fine details are utilized to
technique many multiscales transform techniques have design fusion weights because the strength of fine details
been evolved, such as Laplacian Pyramid [12], Ratio- indicates in-focused regions. Out-of-focus region depicts
of-low-pass Pyramid [13], Morphological Pyramid [12], weak fine texture details. Therefore, an in-focus region
and Gradient Pyramid [14]. After the development of yields stronger texture details (larger fusion weights) than
the wavelet theory [15], the image fusion algorithms are the out-of-focus region.
proposed to merge multiple multifocus images together.
However, the wavelet transform cannot preserve salient 2.1 Computation of Weight Map Function
features in source images and leads to some artifacts and
inconsistency [16]. Recently, various multifocus image The two-scale decomposition based on BLT helps to sep-
fusion approaches have been developed [13–20]. Image arate coarser details (called the base layer) from finer
fusion, using a multiresolution approach, can be per- details (called the detail layer). The base layer is seg-
formed efficiently through the use of the local image regated into different spatial bands called levels. New
contrast level [16]. It has been verified that the perfor- spatial bands pertaining to the resultant fused image
mance of multifocus image fusion approaches is remark- are obtained appropriately by combining these spatial
ably affected by the selection of fusion rules, which are bands based on weight map function at each level. This
based on blurriness measures [21]. These image fusion technique is claimed to be suitable for utilizing the mul-
techniques are unable to preserve edges and weak tex- tiresolution technique for multifocus image fusion as it
ture details simultaneously [22]. The goal of this paper does produce a fused image with better texture and edge
is to extend this multiresolution method to the prob- details without introducing visible artifacts.
lem of multifocus image fusion by utilizing a BLT that
employs a product of a Gaussian kernel in the spatial Consider two input images with the same spatial reso-
domain and a Gaussian kernel in the intensity [22–25]. It lution I 1 and I 2 having dimensions (N × M) pixels. B1
is an edge-preserving filter, which smoothes out only the and B2 are the base layers computed from I 1 and I 2,
fine textures present in the image. In this work, the pyra- respectively. These base layers are computed using BLT
mid approach and two-scale decomposition to extract [17], which is known as an edge-preserving filter:
distinct features based on BLT is proposed for multifo- B1 = BLT(I1 ), (1)
cus image fusion. The unique features and information
are extracted from all source images using BLT; there- B2 = BLT(I2 ). (2)
fore, act as a weight map for the present multifocus image The detailed description of BLT is given in the forthcom-
fusion approach. Moreover, the extracted features across ing section.
source images are utilized to enhance texture details in
the final fused image. This paper throws light on the The weight map function is computed for each input
fusion of multifocus images using BLT filter and pyra- image by subtracting the input image and the corre-
mid decomposition. Therefore the fusion is performed in sponding base layer. This new layer is recognized as a
the transform domain that avoids any artifacts during the detail layer. Let D1 and D2 are the detail layers computed
selection of in-focus regions across input images. from I 1 and I 2 , respectively:

The structure of this paper is illustrated as follows: D1 = I1 − B1 , (3)


Section 1 provides an introduction to image fusion and D2 = I2 − B2 . (4)
pyramid-based algorithms for multifocus fusion. Section
2 describes the proposed approach for multifocus image If the sum of all the weights at any given spatial loca-
fusion. Section 3 deals with discussion and analysis of tion equals unity, then such weights are called normalized
experimental results and comparative performance anal- weights, the fusion weights (w1 = D1 and w2 = D2 ) are
ysis of the proposed technique with existing state-of-the normalized.
art technique and finally Section 4 concludes the paper.
The resultant fused base layer B is calculated as

2. PROPOSED APPROACH B(x, y) = G{w1 (x, y)}L{B1 (x, y)}


+ G{w2 (x, y)}L{B2 (x, y)} (5)
In the first step of the present approach decomposition
of input, images are carried into base layers and detail where G(·) and L(·) are the Gaussian pyramid and
layers. The purpose of two-scale decomposition is to Laplacian pyramid of weight function and base layer,
SIMRANDEEP SINGH ET AL.: MULTI-FOCUS IMAGE FUSION 3

respectively. w1 and w2 are recognized as the weight func- the weight of a pixel and the weight of pixels decreased
tions that are computed from weight layer and base layer with the intensity differences. The smoothing operation
respectively that determine the relative weight given to depends on the Gaussian spatial kernel and Gaussian
the corresponding source images at pixel location (x, y). range kernel g is analyzed to be an edge-stopping func-
The fused base layer (B) is required to be composed of tion for a pixel s and BLT produces the output given by
the constituent base layers B1 and B2 , so the probable Equation (6).
solution is a combination of these images in an additive      
manner. x y I(x , y )gσs (x − x , y − y )gσr
(I(x, y) − I(x − y ))
BLT(x, y) =    
, (6)
In the present method, a pyramid is generated for each x y gσs (x − x , y − y )gσr
base layer through filtering and sub-sampling the prede- (I(x, y) − I(x − y ))
cessor. The reduced version of the input image is gen-  
(x2 +y2 )
erally used for successive images in the pyramid and where gσs (x, y) = exp − 2σ 2
s
because of such representation, this mechanism is also
stated as the multiresolution representation. Details and and
features of an image with coarser approximations are  
â2
represented by the successive levels of an input image gσr (â) = exp ,
2σr2
pyramid representation. The pyramidal structure is com-
puted by the process of convolution, where the current where (x , y ) are the neighborhood pixels of position
approximation of the base layer is convolved with the (x, y) in an image, the extent of the spatial kernel is
Gaussian filter and known as the Gaussian pyramid of denoted by σ s and the minimum amplitude by σ r . Dif-
the base layer. Laplacian pyramid is generated from this ferent values of σ s and σ r affect the filtered image in
structure while performing a sequence of filter-subtract- a different way. The final fused image is computed in
decimate (FSD) operations. The whole mechanism is well Equation (7).
explained in Figure 2. The spatially band-passed images 1 
of the input image are captured by this set of bands. A F(x, y) = B(x, y) + Dk (x, y), k = 1, 2, 3, . . . , N,
N
unique local spatial information about the scene will be k
provided by each band. The pyramidal representations of (7)
all the input base layers are then appropriately combined
at every level using a pre-defined weight function based where N denotes the number of detail layers.
on the texture details provided by BLT filter. A modi-
fied Laplacian pyramid will be generated by combining 2.3 Pyramid Decomposition and Image Fusion
this process, the fusion of the base layer at that partic-
The Gaussian pyramid contains a sequel arrangement of
ular level will be represented by the fusion of the base
images where every next image is a copy of a low-pass fil-
layers. A reverse transformation is applied on the fused
tered image of its previous image [5]. A Gaussian image
base layer and the final resultant base layer can then be
pyramid is mostly preferred in most of the multiresolu-
reconstructed. The detailed description of the computa-
tion pyramids for image processing. Fast access to image
tion of the Laplacian pyramid and Gaussian pyramid is
features of specific frequencies or sizes is easily obtained
given in Section 2.3.
through this. It stores a set of successively low-pass fil-
tered and subsample versions of an image. A local, sym-
2.2 Bilateral Filter and Detail Layer Fusion metric weighting function resembling Gaussian proba-
bility distribution is used. The process of averaging is
Bilateral filtering is a recently proposed edge-preserving
carried out by a convolving that weigh the function and
approach. The output of this filter is the weighted aver-
image. The sequence of images is produced, named G0 ,
age of the input without smoothing edges and also it is
G1 , G2 , . . . , GN and a pyramid generated through these
a nonlinear filter. Size and contrast are two important
images is called the Gaussian pyramid, as computed in
parameters of the image, so the features to preserve these
Equation (8):
parameters are involved in it. It can be applied in a non-
iterative manner, so the effect of these parameters is not 
2 
2

cumulative. Gl (i, j) = w(m, n)Gl−1 (2i + m, 2j + n). (8)


m=−2 n=−2

Consider standard Gaussian filter having spatial kernel Window size is selected as 5-by-5 elements as suggested
f and intensity domain function g, which contributes to by Burt et al. [6] because “it provides adequate filtering at
4 SIMRANDEEP SINGH ET AL.: MULTI-FOCUS IMAGE FUSION

low computation cost”. To choose a kernel denoted by w information for each band is preserved by this pyramid.
(m, n), which is a weighting pattern, certain constraints A compact two-dimensional filter is used to create a pyra-
need to be considered and are listed as in equations (9), mid by low pass-filtering an image to ground level. Every
(10), (11) and (12). other pixel is removed in all the rows and columns to
obtain a subsampled image from the filtered image G0
w(m, n) = ŵ(m).ŵ(n). (9) and new reduced subsampled image is denoted G1 . A
Gaussian pyramid is obtained in this process after under-
It is one-dimensional, having length (−2, −1, 0, 1, 2) that
going many iterations named G0 , G1 , G2 , G3 . . . Gn [26].
is 5, function ŵ is normalized and symmetric.


2 
0
ŵ (m) = 1, (10) Gk (i, j) = Gk−1 [(2i + m, 2j + n)], (13)
m n
m=−2

ŵ(i) = ŵ(−i) for i = 0, 1, 2. (11) where k = 1, 2, 3, . . . , N.

One more critical constraint called equal contribution is The next extension of the Gaussian pyramid is Lapla-
also involved. This ensures total weight contributed to cian pyramid [12,14,27,28], which is generated using a
the next higher nodes should be the same, i.e. ( = 1/4). set of band-passed images. In this pyramid, each next
Let ŵ(0) = a, ŵ(−1) = ŵ(1) = b, and ŵ(−2) = ŵ(2) = image is obtained by filtering a previous image through a
ccondition for equal contribution is a + 2c = 2b. All the band-pass filter. Band-passed images are acquired in the
constraints mentioned above can be satisfied while con- Gaussian pyramid by computing the difference between
sidering the weights given in Equation (12). successive levels of low-pass images [3,29]. So for the
Laplacian pyramid, Gaussian pyramid becomes the basic
ŵ(0) = a building block. In the proposed approach, the Laplacian
1 pyramid and Gaussian pyramid for each input image and

w(−1) = ŵ(1) =
4 corresponding weight maps are computed. A strength
ŵ(−1) = ŵ(1) = 1/4 − a/2. (12) measure is required through which a pixel with the best
information for a specific location can be chosen for a
Figure 1 illustrates the schematic diagram for the pro- sample. Expand G1 to the same size as G0 . Once G1
posed approach using pyramid representation. The input becomes comparable to G0 both may be subtracted and
image sequence of Mandrill and weight maps computed resultant image is band-passed image denoted by L0 and
based on texture analysis (i.e. detail layers) are illustrated so on. The expanded image Gk,1 is given by Equation
in Figure 2. The pyramid represents an image as a sum- (15). To calculate the prediction error L0 (ij), the original
mation of spatially band-passed images. Local spatial image G0 (ij) is subtracted from and G1 (ij), as denoted in

Figure 1: The schematic diagram of the proposed fusion approach


SIMRANDEEP SINGH ET AL.: MULTI-FOCUS IMAGE FUSION 5

Figure 2: The proposed fusion approach for two source images

Equation (14). Where G1 (ij) is the resultant image cal-


culated after passing G0 through suitable low-pass filters
[30].

L0 (ij) = G0 (ij)—G1 (ij)


L2 (ij) = G1 (ij)—G2 (ij), (14)

Gk.l (i, j) = 4 Gk,l−1
m n

× [(2i + m)/2, (2j + n)/2]f (m, n). (15)

G1 is expanded to the same size as of G0 , G0 and G1


are subtracted to produce a new band-passed Lapla-
cian image L0 and new Laplacian pyramid having band-
passed images of decreasing size and spatial frequency
denoted by L0 , L1 , L2 , . . . , Ln−1 , can be built according
to Equation (14)

Lk = Gk − G∗k+1 , k = 0, N − 1. (16)

G∗k+1 is the expanded version of Gk+1 , whose size is the


same as that of Gk . A well-focused image is generated by
adding a modified Laplacian pyramid image with a fused
detail layer.
Figure 3: Image reconstruction from modified Laplacian pyra-
mid of the base layers. Please note that the fused detail layer is
The Laplacian pyramid of the base layer Lk multiplied combined with the fused base layer (B) to obtain final all in-focus
with the corresponding Gaussian pyramid of the detail fused image (F)
6 SIMRANDEEP SINGH ET AL.: MULTI-FOCUS IMAGE FUSION

layer and summing over k yields modified Laplacian different images. Five multifocus image sequences were
pyramid LPmk , where m represents the number of input used for experimentation.
images. Modified Laplacian pyramids LPm k are added by

expanding each level, as shown in Figure 3. In most of the cases, the proposed method produces sat-
isfactory results both visually and quantitatively. Figures
4–8 show the comparison of the proposed algorithm
3. DISCUSSION AND ANALYSIS OF
with ground truth and six other state-of-the-art multi-
EXPERIMENTAL RESULTS
focus image fusion methods on multiple data sets. Out
The proposed method has experimented on a vari- of these few samples the Mandrill, Circle–Triangle, Folk,
ety of multifocused image sets (including synthetic and Character, and Book image sequences, have been dis-
real images) [31]. Character (256 × 256), Circle-Triangle cussed here. Well-focused pixels in an input sequence
(290 × 290), and Mandrill (500 × 480) are Synthetic yield strong texture details that produce higher weights
image sets. In Character image set, focus is on the objects than blurry pixels. Therefore, we have introduced our
of different shapes and sizes, in the Circle–Triangle new measures to generate adaptive weights for mul-
images focus is alternatively on Circle and triangle, tifocus image fusion. It may be noted from Figure 2
whereas Mandrill images divide focused regions horizon- that adaptive weight-based modified bands of each input
tally [31]. The established real image sets taken from mul- image have been initially fused into the single fused
tifocus camera are Folk (512 × 512) with ground truth base layer, which is then combined with the fused detail
and Book (1280 × 960) without ground truth. The Folk layer to obtain the final composite image. Therefore,
and Book image sets have different objects focused on this model preserves strong edges and texture details in

Figure 4: (a) Ground truth (b) and (c) Source images (d) Proposed (e) DWT (f) SIDWT (g) NVF (h) SPF (i) MSMFM (j) SDMF

Figure 5: (a) Ground truth (b) and (c) Source images (d) Proposed (e) DWT (f) SIDWT (g) NVF (h) SPF (i) MSMFM (j) SDMF
SIMRANDEEP SINGH ET AL.: MULTI-FOCUS IMAGE FUSION 7

Figure 6: (a) Ground truth (b) and (c) Source images (d) Proposed (e) DWT (f) SIDWT (g) NVF (h) SPF (i) MSMFM (j) SDMF

Figure 7: (a) Ground truth (b) and (c) Source images (d) Proposed (e) DWT (f) SIDWT (g) NVF (h) SPF (i) MSMFM (j) SDMF

multifocus image fusion problem using the conventional Figures 4–8 depict the comparison of proposed fusion
pyramid approach, which is performed by applying dif- results with others. Images illustrated in Figures 4(e),
ferent gain-control factors on the base layer computed 5(e), 6(e), 7(e), and 8(e) are produced using the
from BLT to each band-passed component. method of Discrete Wavelet Transform (DWT) [28],

Figure 8: (a) and (b) Source images (c) Proposed (d) DWT (e) SIDWT (f) NVF (g) SPF (h) MSMFM (i) SDMF
8 SIMRANDEEP SINGH ET AL.: MULTI-FOCUS IMAGE FUSION

Figures 4(f), 5(f), 6(f), 7(f), and 8 (f) are produced evaluated based on the performance of statistical eval-
using the method of Shift-invariant Discrete Wavelet uation metrics, such as Entropy [36], Mean, Standard
Transformation (SIDWT) [32], Figures 4(g), 5(g), 6(g), deviation [37], Average gradient [38], and PSNR [36].
7(g), and 8(g) are produced using the method of Noise The average amount of information conveyed in an image
Visibility Function (NVF) [33], and Figures 4(h), 5(h), referred to as entropy. The proximity of the entropy val-
6(h), 7(h), and 8(h) are produced using the method ues illustrates better fusion. If a reference image (ground
of Salience Preserving Fusion SPF [22]. Figures 4(i), truth) is unavailable, desirable values for entropy are
5(i), 6(i), 7(i), and 8(i) are produced through multi- higher for better fusion results. Entropy may be defined
scale morphological focus-measure MSMFM [34] and as Equation (17).
Figures 4(j), 5(j), 6(j), 7(j), and 8(j) are generated
using saliency detection-based multifocus image fusion L −1
2
(SDMF) [35]. EI = − p(sj )log2 (p(sj )), (17)
j=1
Images shown in Figures 4(d), 5(d), 6(d), 7(d), 8(d),
and 9(d) are generated by the proposed method. The where L is the number of gray levels, p(sj ) is the probabil-
results produced by DWT, SIDWT, NVF, and SPF lose ity of occurrence of gray level sj in an image I.
strong edges and weak texture details. The results of
DWT, SIDWT, NVF, and SPF in all experiments are The second metric, Mean, measures the central tendency
implemented and compared. The edge details present the of the given data, dispersion around means is referred to
results of SIDWT (Figures 4f, 5f, 6f, 7f, and 8f) are com- as standard deviation. A higher value of mean and stan-
parable to those of the proposed algorithm. The result dard deviation is expected for a fused image to have better
of NVF (Figures 4g, 5g, 6g, 7g, and 8g) and SPF (Fig- fusion results. For an image I of size M × N mean and
ures 4h, 5h, 6h, 7h, and 8h) suffers a little blurry. The standard deviation are given by Equations (18) and (19),
result of DWT (Figures 4e, 5e, 6e, 7e, and 8e) shows respectively, where μI defines mean and σI symbolizes
serious halo artifacts on object boundaries. Strong edges standard deviation:
are preserved in the proposed algorithm (Figures 4d,
5d, 6d, 7d, and 8d), yet at the same time, fine texture 
M 
N
details are well preserved. From the present fused image μI = I(i, j) (18)
of the proposed method shown in (Figures 4d, 5d, 6d, i=1 j=1
7d, and 8d), it should be noted that both the spatial res-
olutions and clarity have been bettered than other tech- and
niques. Therefore, from a subjective evaluation point of
⎡ ⎤1
view, the overall effectiveness of the proposed approach is 
M 
N 2
1
better. σI = ⎣ (I(i, j) − μI )2 ⎦ . (19)
MN − 1 i=1 j=1
Qualitative analysis alone is not adequate to check the
efficiency of fusion algorithm; rather, certain quantita- For image I I,x is a gradient in the horizontal direction
tive analysis is required. Proposed fusion algorithm is and I,y is gradients in the vertical direction the average

Figure 9: The proposed fusion results for “Clock” grayscale images. (a), (b) The original source images, (c) The Image reconstructed by
the present approach.
SIMRANDEEP SINGH ET AL.: MULTI-FOCUS IMAGE FUSION 9

gradient denoted by I is given by Equation (20): [19] value decreases, match also will decrease. Structural
1
similarity (SSIM) is an indicative of the perceptual struc-
1   (I,x − I,y ) 2
M N 2 2 tural similarity between the ground truth image IG and
I = . (20) the fused image IF . Mathematically SSIM map between
MN 2
i=1 j=1 IF andIG is defined as follows [39]:
The values of the average gradient for ground truth and σIFIG 2μIF μIG 2σIF σIG
SSIMmap (IF , IG ) = , (22)
fused image should be closer for the better fused results. σIG σIF μ2IF + μ2IG σI2F + σI2G
If ground truth is not available, a higher value of aver-
age gradient is expected for better fusion results. PSNR where σIFIG is the blockwise covariance between IF and IG .
is peak signal to noise ratio and it is a most important Whereas μIF and σIF are the blockwise mean and stan-
parameter in image quality analysis, it indicates similar- dard deviation of the image IF , respectively. The mean
ity between two images. The PSNR between the fused of the SSIMmap is the value of SSIM between IF and IG .
image IF and the ground truth IG , each of size M × N, For a perfect similarity and dissimilarity between the two
is calculated as in Equation (21): images, SSIM equals 1 and 0. PSNR and SSIM can be cal-
culated for only those data sets, where ground truth is
 available. For book image set, these two values are not
2552
PSNR = 10ln 1 M N . mentioned in Table 1 due to the unavailability of ground
j=1 (IF (i, j) − IG (i, j))
2
MN i=1 truth. Table 1 shows the results of the quantitative evalua-
(21) tion results by measuring six evaluation metrics. The pro-
posed method achieves superior results for most cases.
Higher PSNR value results in more similarity between The values of the fusion results of the four image set are
the two images. Ideal PSNR value is infinity (∞) in the not entirely consistent because the input image sequences
case of a perfect match between the two images. If PSNR contain abundant texture information. The fusion metric

Table 1: Comparison analysis of the proposed method, DWT, SIDWT, NCF, and SPF based on different metrics
Image name Method name Entropy Mean Standard deviation Average gradient PSNR SSIM
Mandrill Proposed 7.3899 103.7832 42.4228 13.1304 38.0121 0.9867
DWT 7.3522 103.2863 43.5048 13.0843 37.1776 0.9132
SIDWT 7.2930 103.5147 43.0959 12.7465 36.5299 0.9338
NVF 6.7511 102.5051 43.9820 13.1152 40.2021 0.9726
SPF 7.2957 102.6710 40.9471 11.2781 30.0283 0.9621
MSMFM 3.7725 103.1664 41.9002 2.3937 5.6469 0.9202
SDMF 3.8845 102.4567 41.8737 4.0757 5.5786 0.9017
Folk Proposed 7.0198 60.6066 55.2614 10.5477 38.2031 0.9864
DWT 6.7214 60.0398 49.7508 4.4169 36.6637 0.9713
SIDWT 6.6712 60.0715 49.5527 4.1550 38.1985 0.9792
NVF 6.6693 60.3855 49.2456 4.2898 34.6754 0.9574
SPF 6.6982 59.6082 48.3021 3.7079 31.775 0.9657
MSMFM 3.5524 60.2998 55.2513 1.5540 4.1730 0.8972
SDMF 3.7122 59.4587 101.7007 2.7543 4.0924 0.9388
Character Proposed 6.1233 204.7794 54.0146 9.0422 23.8489 0.9776
DWT 6.0187 204.4417 60.7086 10.8270 23.2796 0.8581
SIDWT 5.6403 205.1699 60.0636 9.8801 26.3396 0.9005
NVF 5.1212 199.9718 58.7131 13.2725 19.7049 0.8123
SPF 5.7247 200.9493 54.5767 8.99981 7.3096 0.8312
MSMFM 2.8498 204.6552 46.1469 1.8094 10.7701 0.8005
SDMF 3.1885 204.7112 48.9347 2.4592 10.6738 0.8671
Circle-triangle Proposed 2.5039 90.0598 47.0569 1.1608 43.9778 0.9934
DWT 2.3281 89.9076 46.5003 0.8087 43.8563 0.9776
SIDWT 1.9494 89.9061 46.4538 0.6810 42.7015 0.9885
NVF 1.7513 89.8816 46.1118 0.8527 37.2068 0.9726
SPF 2.4403 88.3151 46.6937 1.2824 28.9115 0.9621
MSMFM 1.8775 90.0346 47.0168 0.4366 5.2467 0.9495
SDMF 2.3509 89.9856 86.5930 0.8721 5.1928 0.9492
Book Proposed 7.3593 84.9278 63.5783 5.4834 – –
DWT 7.3492 84.9136 62.4994 5.3162 – –
SIDWT 7.2957 84.9093 62.2903 4.9833 – –
NVF 7.3187 85.5550 61.6245 5.4877 – –
SPF 7.3306 84.5197 60.0008 3.9981 – –
MSMFM 7.2589 84.9234 62.3461 4.3628 – –
SDMF 7.15365 84.6345 59.6301 5.3754 – –
10 SIMRANDEEP SINGH ET AL.: MULTI-FOCUS IMAGE FUSION

values of the proposed fusion results for Folk and Cir- 4. H. S. Singh, and N. S. Simrandeep, “Multi-resolution rep-
cle–Triangle data set are greater than those of other meth- resentation of multifocus image fusion using Gaussian and
ods. The results indicate that for Mandrill, Folk, Char- Laplacian pyramids,” Int. J. Adv. Res. Comput. Sci. Softw.
Eng., Vol. 3, no. 11, pp. 1639–1642, 2013.
acter, and Circle–Triangle data set, the present method
leads to the best performance for Entropy and Mean. 5. K. Kazemi, and H. A. Moghaddam, “Fusion of multifo-
Standard deviation, Average gradient, and PSNR are also cus images using discrete multiwavelet transform,” IEEE
optimal in most of the cases. Default parameter settings Int. Conf. Multisens. Fusion Integr. Intell. Syst, pp. 167–172,
for two-scale decomposition BLT is considered as 2 and 2003.
0.8 for spatial kernel σs and the minimum amplitude σr ,
6. S.-G. Huang, “Wavelet for image fusion.” Graduate Institute
respectively [17]. Based on the analysis of experimen- of Communication Engineering and Department of Electri-
tal results, it has been verified that the present approach cal Engineering, National Taiwan University, 2010.
yields a better result, in both cases visually and quantita-
tively for multifocus image fusion. The usefulness of the 7. V. D. Maitreyi Abhyankar, and A. Khaparde, “Spatial
pyramid approach and edge-preserving filter for multi- domain decision based image fusion using superimposition,”
in 2016 IEEE/ACIS 15th Int. Conf. Comput. Inf. Sci., 2016,
focus image fusion with texture detail enhancement has pp. 1–6.
been demonstrated in this algorithm.
8. Q. Zhang, Y. Liu, R. S. Blum, J. Han, and D. Tao, “Sparse
representation based multi-sensor image fusion for multi-
4. CONCLUSION AND FUTURE ASPECTS focus and multi-modality images: A review,” Inf. Fusion,
Vol. 40, pp. 57–75, 2018.
The proposed algorithm describes a simple and robust
method for multifocus image fusion, which makes it pos- 9. H. Singh, V. Kumar, and S. Bhooshan, “A novel approach
sible to produce an all-in-focus image containing strong for detail-enhanced exposure fusion using guided filter,”
Sci. World J, Vol. 2014, pp. 1–8, 2014.
edges and texture details using a conventional pyramid
approach. The weight maps are computed based on the 10. Q. Jiang, X. Jin, S. J. Lee, and S. Yao, “A novel multi-focus
texture analysis using BLT filter. The proposed method image fusion method based on stationary wavelet trans-
modifies band-passed components of base layers across form and local features of fuzzy sets,” IEEE. Access., Vol.
input images in the transform domain and then con- 5, pp. 20286–20302, 2017.
structs a single fused image by adding fine details. Pre-
11. I. De, and B. Chanda, “A simple and efficient algorithm for
liminary results show promise in edge and texture details multifocus image fusion using morphological wavelets,”
manipulation independently in the fused image. Differ- Signal. Process., Vol. 86, no. 5, pp. 924–936, 2006.
ent datasets are experimented to test the performance
of our method. Experimental results show that it pro- 12. S. Krishnamoorthy, and K. P. Soman, “Implementation
duces an all-in-focus image by selecting the in-focus and comparative study of image fusion algorithms,” Int. J.
Comput. Appl, Vol. 9, no. 2, pp. 25–35, 2010.
region across source images. It is worth concentrating on
many different thrilling possible applications of the con- 13. K. Barlos, et al., “Image fusion by ratio of low pass filter,”
struction of a fused image using the proposed approach Tetrahedron Lett., Vol. 30, no. 30, pp. 3943–3946, 1989.
in future. However, to improve fusion performance the
present approach needs to be investigated further using 14. W.-W. Wang, P.-L. Shui, and G.-X. Song. “Multifocus image
optimization techniques. fusion in wavelet domain,” pp. 2887–2890, 2004.

15. Y. Dai, Z. Zhou, and L. Xu, “The application of multi-


modality medical image fusion based method to cerebral
REFERENCES
infarction,” Eurasip J. Image Video Process, Vol. 2017, no. 1,
1. S. Gautam, S. Singh, and N. Ramanpreet. “A centroid pp. 55, 2017.
matching approach for real time foreground segmentation
and occlusion handling,” Proc. 2015 1st Int. Conf. Next 16. Q. Wang, and Y. Shen, “The effects of fusion structures
Gener. Comput. Technol. NGCT 2015, September, 2016, pp. on image fusion performances,” Conf. Rec. IEEE Instrum.
663–667. Meas. Technol. Conf, Vol. 1, no. May, pp. 468–471, 2004.

2. F. Sadjadi, “Comparative image fusion analysais.” 2005 17. H. Singh, V. Kumar, and S. Bhooshan, “Anisotropic dif-
IEEE Computer Society Conference on Computer Vision and fusion for details enhancement in multiexposure image
Pattern Recognition (CVPR’05)-Workshops. IEEE, 2005. fusion,” ISRN Signal Process, Vol. 2013, pp. 1–18, 2013.

3. W. Wang, and F. Chang, “A multi-focus image fusion 18. D. P. Bavirisetti, and R. Dhuli, “Two-scale image fusion
method based on Laplacian pyramid,” J. Comput., Vol. 6, of visible and infrared images using saliency detection,”
no. 12, pp. 2559–2566, 2011. Infrared Phys. Technol, Vol. 76, pp. 52–64, 2016.
SIMRANDEEP SINGH ET AL.: MULTI-FOCUS IMAGE FUSION 11

19. R. Kaur, and S. Singh. “An artificial neural network based 30. Y. Song, M. Li, Q. Li, and L. Sun, “A new wavelet based
approach to calculate BER in CDMA for multiuser detec- multi-focus image fusion scheme and its application on opti-
tion using MEM,” Proc. 2016 2nd Int. Conf. Next Gener. cal microscopy,” in IEEE Int. Conf. Robot. Biomimetics,
Comput. Technol. NGCT 2016, no. October, pp. 450–455, ROBIO 2006, pp. 401–405.
2017.
31. A. Saha, G. Bhatnagar, and Q. M. J. Wu, “Mutual spec-
20. S. Li, X. Kang, J. Hu, and B. Yang, “Image matting for fusion tral residual approach for multifocus image fusion,”
of multi-focus images in dynamic scenes,” Inf. Fusion, Vol. Digit. Signal Process, Vol. 23, no. 4, pp. 1121–1135,
14, no. 2, pp. 147–162, 2013. 2013.

21. D. Y. Tsai, Y. Lee, and E. Matsuyama, “Information entropy 32. W. Xin, Y. L. Wei, and L. Fu. “A new multi-source image
measure for evaluation of image quality,” J. Digit. Imaging, sequence fusion algorithm based on sidwt,” Proc. 2013 7th
Vol. 21, no. 3, pp. 338–347, 2008. Int. Conf. Image Graph. ICIG 2013, no. 3, pp. 568–571,
2013.
22. C. Pal, A. Chakrabarti, and R. Ghosh, “A brief survey of
recent edge-preserving smoothing algorithms on digital 33. G. BHATNAGAR, and Q. M. J. WU, “An image fusion
images,” arXiv preprint arXiv:1503.07297, 2015. framework based on human visual system in Framelet
domain,” Int. J. Wavelets, Multiresolution Inf. Process., Vol.
23. S. Raman, and S. Chaudhuri, “Bilateral Filter Based Com- 10, no. 01, pp. 1250002, 2011.
positing for Variable Exposure Photography.” in Euro-
graphics (short papers), pp. 1–4, 2009. 34. Y. Zhang, X. Bai, and T. Wang, “Boundary finding based
multi-focus image fusion through multi-scale morpho-
24. Bob Fisher. “Bilateral Filtering,” 2004. logical focus-measure,” Inf. Fusion, Vol. 35, pp. 81–101,
2017.
25. J. J. Francis, and G. De Jager, “The bilateral median filter,”
SAIEE Africa Res. J, Vol. 96, no. 2, pp. 106–111, 2005. 35. D. P. Bavirisetti, and R. Dhuli, “Multi-focus image fusion
using multi-scale image decomposition and saliency detec-
26. S. Liu, J. Zhang, and J. Chen. “Multi-focus image fusion tion,” Ain Shams Eng. J., Vol. 9, no. 4, pp. 1103–1117,
using Gaussian filter and dynamic programming,” Proc. 2018.
9th Asia-Pacific Signal Inf. Process. Assoc. Annu. Summit
Conf. APSIPA ASC 2017, vol. 2018-Febru, no. December, 36. P. Sreeja, and S. Hariharan, “An improved feature based
pp. 1182–1185, 2018. image fusion technique for enhancement of liver lesions,”
Biocybern. Biomed. Eng, Vol. 38, no. 3, pp. 611–623, 2018.
27. A. Dogra, B. Goyal, and S. Agrawal, “Osseous and dig-
ital subtraction angiography image fusion via various 37. M. Manchanda, and R. Sharma, “An improved multi-
enhancement schemes and Laplacian pyramid transforma- modal medical image fusion algorithm based on fuzzy
tions,” Futur. Gener. Comput. Syst, Vol. 82, pp. 149–157, transform,” J. Vis. Commun. Image Represent, Vol. 51, no.
2018. December 2016, pp. 76–94, 2018.

28. N. Indhumadhi, and G. Padmavathi, “Enhanced image 38. A. Gosain, and J. Singh, “Proceedings of the 3rd inter-
fusion algorithm using Laplacian pyramid and spatial fre- national conference on frontiers of intelligent comput-
quency based wavelet algorithm,” Int. J. Comput. Sci. Eng, ing: theory and applications (FICTA) 2014. Vol. 327, pp.
Vol. 1, no. 5, pp. 298–303, 2011. 305–316, 2015.

29. S. Paris, S. W. Hasinoff, and J. Kautz, “Local Laplacian 39. X. Jin, Q. Jiang, S. Yao, and K. He, “A survey of infrared and
filters: edge-aware image processing with a Laplacian pyra- visual image fusion methods,” Infrared Phys. Technol., Vol.
mid,” Siggraph, Vol. 58, no. 3, pp. 81–91, 2011. 85, pp. 478–501, 2017.
12 SIMRANDEEP SINGH ET AL.: MULTI-FOCUS IMAGE FUSION

Authors Nitin Mittal received his B.Tech. and


M.Tech. degrees in ECE from Kurukshetra
Simrandeep Singh received his B.Tech. University, Kurukshetra, India, in 2006
and M.Tech. degrees in Electronics and and 2009, respectively. He received his
Communication Engineering (ECE) from Ph.D. in ECE from Chandigarh Univer-
the Punjab Technical University in 2008 sity, Mohali, India, in 2017. He is working
and 2015, respectively. He is pursu- as Associate Professor in the Department
ing Ph.D. from Chandigarh University, of ECE, Chandigarh University, Mohali,
Gharuan, Punjab, India. Currently, he is India. He has 10 years of teaching experience. He has more
working as an Assistant Professor in the than 50 research publications in various fields. His research
Department of ECE, Chandigarh University, Gharuan, India. interests include Wireless Sensor Networks, Digital Signal Pro-
He has 11 years of teaching experience. He has more than 15 cessing, Multifocus Image Fusion, Exposure Fusion, and Edge-
research publications in various fields. His current research Preserving Filters.
interests include diatom analysis based on multifocus image
fusion, exposure fusion, edge-preserving filters, Wireless Sen-
Email: mittal.nitin84@gmail.com
sor Networks, and Digital Signal Processing.
Harbinder Singh has done Ph.D. degree
Corresponding author. Email: simrandeepsingh.cgc@gmail. in ECE at the Jaypee University of Infor-
com mation Technology, Waknaghat, India,
received the M.Tech. degree in ECE from
Punjab Technical University, Punjab, the
B.Tech. degree in ECE from Himachal
Pradesh University. He is the author or co-
author of 24 papers published in national,
international conferences and journals. His current research
interests include diatom analysis based on multifocus image
fusion, exposure fusion, edge-preserving filter.

Email: harbinder.ece@gmail.com

You might also like