Sei sulla pagina 1di 14

Information Fusion 22 (2015) 105–118

Contents lists available at ScienceDirect

Information Fusion
journal homepage: www.elsevier.com/locate/inffus

Quadtree-based multi-focus image fusion using a weighted


focus-measure
Xiangzhi Bai a,b,⇑, Yu Zhang a, Fugen Zhou a, Bindang Xue a
a
Image Processing Center, Beijing University of Aeronautics and Astronautics, Beijing 100191, China
b
State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing 100191, China

a r t i c l e i n f o a b s t r a c t

Article history: The purpose of multi-focus image fusion is integrating the partially focused images into one single image
Received 12 November 2013 which is focused everywhere. To achieve this purpose, we propose a new quadtree-based algorithm for
Received in revised form 26 February 2014 multi-focus image fusion. In this work, an effective quadtree decomposition strategy is presented.
Accepted 21 May 2014
According to the proposed decomposition strategy, the source images are decomposed into blocks with
Available online 6 June 2014
optimal sizes in a quadtree structure. And in this tree structure, the focused regions are detected by using
a new weighted focus-measure, named as the sum of the weighted modified Laplacian. Finally, the
Keywords:
focused regions could be well extracted from the source images and reconstructed to produce one fully
Multi-focus image fusion
Quadtree decomposition strategy
focused image. Moreover, the new weighted focus-measure performs better than the commonly used
Quadtree structure focus-measures on the detection of the focused regions, since it is sensitive to the homogeneous regions.
Weighted focus-measure The proposed algorithm is simple yet effective, because of the quadtree decomposition strategy and the
Sum of the weighted modified Laplacian new weighted focus-measure. To do the comparison, the proposed algorithm is compared with several
existing fusion algorithms, in both the qualitative and quantitative ways. The experimental results show
that the proposed algorithm yields good results.
Ó 2014 Elsevier B.V. All rights reserved.

1. Introduction are processed according to some criteria and transformed back to


produce the fusion image. However, the transform domain algo-
It is known that the lens of a camera can only focus at one dis- rithms may encounter some conditions when dealing with the high
tance or on a plane at a time [1,2]. Thereby, only the objects falling dimensional signals like images [2]. For example, the transform
in the depth of field (DOF) of the lens appear to be sharp, while the domain algorithms usually generate the fusion image globally,
objects out of the DOF are blurred [1–3]. As a result, the multiple thus their fusion image would be easily resulted in deterioration
objects of interest may not be all focused in one single image. To due to misregistration [7,9].
produce one image in which all objects are in focus, image fusion The spatial domain algorithms are based on a local operation.
is a kind of method. As is known, the image fusion techniques have They can avoid the global problems and preserve more original
been widely used in many applications, such as medical image information from the source images than the transform domain
fusion [4], infrared and visual image fusion [5] and remote sensing algorithms. In general, spatial domain algorithms can be classified
image fusion [6]. The term fusion for the multi-focus images is as: pixel-based, block-based, and region-based [3]. The pixel-
integrating the multiple images of the same scene to form one fully based fusion algorithms just consider the single pixels or use
focused image, which is suitable for visual perception or computer the information in the local neighborhoods, which could lead to
processing [1–3,7,8]. some undesirable side effects [2], such as reduced contrast, and
In the past decades, different genres of multi-focus image fusion blocking artifacts. To resolve this problem, some block-based algo-
algorithms have been proposed. Most of these algorithms are done rithms [8–10] have been proposed. In the block-based algorithms,
either in the spatial domain or transform domain. the source images are firstly decomposed into blocks with equal
The transform domain algorithms usually extract coefficients size. Then the focused blocks can be detected by measuring the
with respect to some localized bases [2]. Then, these coefficients focus on the corresponding blocks. However, a large block may
contain both the focused and defocused regions and the focus of
⇑ Corresponding author at: Image Processing Center, Beijing University of a small block could not be well measured [3]. Therefore, the
Aeronautics and Astronautics, Beijing 100191, China. performance of these algorithms is restricted by the selected
E-mail address: jackybxz@buaa.edu.cn (X. Bai). block-size.

http://dx.doi.org/10.1016/j.inffus.2014.05.003
1566-2535/Ó 2014 Elsevier B.V. All rights reserved.
106 X. Bai et al. / Information Fusion 22 (2015) 105–118

In recent years, several more reasonable fusion algorithms have the normalized difference in focus-measures (NDFM) and a con-
been proposed. In [11], Aslantas et al. gave a method to choose the stant multiplier (M), as Eqs. (1) and (2) [3]. At the beginning, each
optimal block-size. Since the optimization process is done recur- source image is subdivided into four blocks. In this paper, the cor-
sively, this algorithm is time-consuming. In [1,12,13], several responding blocks are called as a block-pair (two blocks) or a
region-based algorithms are proposed for multi-focus image fusion. block-set (more than two blocks). Then, the NDFM for each
At first, the source images are segmented into regions, and then the block-pair is compared with the corresponding threshold T, which
focused regions could be found out by comparing the focus-mea- equals to the product of the NDFM calculated from its parent block-
sures of the corresponding regions. However, the region-based pair and M. If NDFM > T, the current block-pair would not be sub-
algorithms are time-consuming and their performance greatly divided again. Otherwise, each block in the current pair should
relies on its previous segmentation results. In [3], De et al. proposed be further subdivided into four smaller blocks. However, in some
a quadtree-based algorithm for multi-focus image fusion. In this cases, the source images could not be subdivided properly, such
algorithm, the source images are firstly decomposed into blocks as the example shown in Fig. 2. In Fig. 2, it is assumed that the
with different sizes in a quadtree structure. In this tree structure, source images are centrosymmetric, and the gradients of the
the focused blocks are detected by measuring the focus on the cor- source images are distributed uniformly. Since the source images
responding blocks. But in some cases, their quadtree decomposition are centrosymmetric, the NDFMs for the corresponding blocks are
method might be not effective, so that the optimal subdivision of equal. Then, the constant value M and the threshold T can be calcu-
the source images could not be achieved. lated and their values are both zero. Consequently, all the blocks
To improve the quality of the fusion image, a new quadtree- will not be further subdivided, and the optimal subdivision of the
based multi-focus image fusion algorithm is proposed. We make source images could not be achieved. Furthermore, De’s quadtree
use of an effective quadtree decomposition strategy to decompose decomposition method is based on the assumption that the energy
the source images into blocks with optimal sizes in a quadtree distribution of an image is related to the energy distribution of its
structure. In this structure the focused blocks are detected by using subregions. However, the energy distribution of an image is
a new weighted focus-measure, named as the sum of the weighted random, thus their decomposition method may not perform
modified Laplacian. Naturally, the focused blocks are merged to effectively in some cases.
form the focused regions [3]. Then, the focused regions are recon-
maxðFM 1 ;    ; FMm Þ  minðFM 1 ;    ; FM m Þ
structed to verify the consistency by using two sequential filters. NDFM ¼ Pm ð1Þ
Moreover, the proposed algorithm is compared with some existing i¼1 FM i

algorithms qualitatively and quantitatively. The main contribution where FM is the focus-measure, and m is the number of the source
of this paper can be concluded as: (1) proposing an effective quad- images.
tree decomposition strategy and (2) presenting a new weighted P4
1 2  
focus-measure. i¼1 NDFM i
M¼4 1
 std NDFM 21 ;    NDFM 24  100 ð2Þ
NDFM
n
2. Proposed algorithm where NDFM1 is the NDFM for the root block-pair, NDFM2i ;
i ¼ 1 . . . 4g are the NDFMs for the block-pairs at the second level,
The goal of this paper is to generate one fully focused image and std represents the standard deviation operator.
from the multi-focus images. To achieve this goal, a new quad-
tree-based algorithm for multi-focus image fusion is proposed in
2.1.1. An effective quadtree decomposition strategy
this section. The general process of the proposed algorithm and a
To achieve the optimal subdivision of the source images, we
demonstration example are illustrated in Fig. 1. The process of
propose an effective quadtree decomposition strategy in this part.
the proposed algorithm is: (1) decompose the source images into
For convenience, the proposed strategy is described for two source
blocks with optimal sizes, according to an effective quadtree
images and the strategy can be easily extended to three or more
decomposition strategy; (2) find out the focused blocks by measur-
source images.
ing the focus of each block-pair in the tree structure; (3) recon-
At the beginning, two source images represent the root block-
struct the focused regions by using two sequential filters.
pair at the first level of the quadtree structure. If the decomposition
This section is organized as follows. Firstly, the method of
criterion is fulfilled in the root block-pair, then each block in the
detecting the focused regions in a quadtree structure is introduced.
pair is subdivided into four quadrants. And this process is repeated
Secondly, to detect the focused blocks accurately, a new focus-
on the smaller block-pairs, until the decomposition criterion is not
measure, named as the weighted modified of Laplacian, is pro-
fulfilled in all the block-pairs.
posed. Thirdly, the focused regions are reconstructed to produce
Ideally, two situations may occur in a block-pair [3]: (1) one
the fusion image. Finally, the comparison between the proposed
block is fully focused and the other is fully defocused (see Situation
algorithm and the previous work is discussed.
1 in Fig. 3); (2) both of the blocks are partially focused (see Situa-
tion 2 in Fig. 3). Obviously, Situation 1 is the desired situation, in
2.1. Detecting the focused regions which the focused block is fully discriminated from the defocused
block. Thereby, the decomposition criterion for a block-pair can be
Detecting the focused regions from the source images is the simply described as: if the block-pair fits Situation 1, then the fully
most crucial part of the proposed algorithm. Here, we make use focused block can be found out in this pair; otherwise, the block-
of the quadtree structure to achieve this purpose. First, the source pair fits Situation 2, each block in the pair should be further subdi-
images are decomposed into blocks with different sizes in a quad- vided into four quadrants.
tree structure, according to a quadtree decomposition method. To find out the situation, the maximum difference in focus-mea-
Then in the tree structure, the focused blocks can be detected by sures (MDFM) and the sum of the maximum difference in gradients
measuring the focus of the corresponding blocks. To obtain an opti- (SMDG) are firstly calculated on the block-pair. If MDFM P
mal subdivision of the source images, the key is to find an appro- 0:98  SMDG, then the block-pair fits Situation 1, otherwise, this
priate quadtree decomposition method. block-pair fits Situation 2. The definitions and explanations of
In [3], De et al. used a quadtree-based method to detect the MDFM and SMDG are described in detail as follows.
focused regions. To decompose the source images, they defined First, calculate the gradient map of each source image.
X. Bai et al. / Information Fusion 22 (2015) 105–118 107

Fig. 1. The general process of the proposed algorithm and a demonstration example.

Therefore, to measure the focus of the block-pairs, the gradients


of the source images should be firstly calculated. The gradient used
here is the proposed weighted modified gradient (detailed infor-
mation is described in Section 2.2).
Second, find out the maximum gradient map and the minimum
gradient map of the source images, respectively. Their calculations
are mathematically expressed as Eqs. (4) and (5), respectively.

GMmax ðx; yÞ ¼ maxðGM1 ðx; yÞ; GM 2 ðx; yÞÞ; ð4Þ

GMmin ðx; yÞ ¼ minðGM1 ðx; yÞ; GM2 ðx; yÞÞ; ð5Þ

where GMmax and GMmin are the maximum and minimum gradient
Fig. 2. A simple example of De’s detection method.
map of all source images, respectively.
In Eqs. (4) and (5), GM1 and GM2 are the gradients of the source
images, respectively. Because GMmax at each location is chosen as
the greater gradient of GM1 and GM2 at the corresponding location,
GMmax approximates the gradient map of a fully focused image,
with respect to the source images [1]. Similarly, GMmin at each
location is selected as the minimum gradient of GM1 and GM2 at
the corresponding location, then it approximates the gradient
map of a fully defocused image, with respect to the source images.
Third, select out the gradient map, maximum gradient map and
minimum gradient map for the block-pair at the block area B, as
shown in Fig. 4.

gradi ðx; yÞ ¼ GMi ðx; yÞ; i ¼ 1; 2; ð6Þ


Fig. 3. Ideal situations on a block-pair with focus on shaded regions. Situation 1: ðx;yÞ2B
one block is fully focused, the other fully defocused; Situation 2: both are partially
focused.
gradmax ðx; yÞ ¼ GMmax ðx; yÞ; ð7Þ
ðx;yÞ2B
GMi ¼ gradientðIi Þ; i ¼ 1; 2; ð3Þ

gradmin ðx; yÞ ¼ GMmin ðx; yÞ; ð8Þ


where {Ii, i = 1, 2} are the source images; {GMi, i = 1, 2} are the corre- ðx;yÞ2B
sponding gradient maps.
For the multi-focus images, a focused block should preserve where {gradi, i = 1, 2} are the gradient maps of the corresponding
more sharp edges than the corresponding defocused blocks [3,9]. blocks of the source images at the block area B; gradmax is the
In images, the gradients indicate the edge information. Then the maximum gradient map of {gradi, i = 1, 2}; gradmin is the minimum
strength of the gradients could be used as a measure of focus. gradient map of {gradi, i = 1, 2}.
108 X. Bai et al. / Information Fusion 22 (2015) 105–118

of the block-pair. According to the explanation of Eqs. (10) and


(11), if a block-pair fits in Situation 1, FMmax of the block-pair
would equal to the focus-measure of the fully focused block and
obtain the extreme maximum focus-measure of this block. FMmin
of this block-pair would equal to the focus-measure of the fully
defocused block and obtain the extreme minimum focus-measure
of this block. In this case, the MDFM of this block-pair will obtain
the extreme maximum value. Therefore, if MDFM of a block-pair
equals to the difference in the focus-measures of a fully focused
block and the corresponding fully defocused block, then this
block-pair should fit Situation 1, in which there is a fully focused
Fig. 4. A pair of block at the block area B. (X, Y) is the upper left corner, and size of
the block-pair is given by K. block and a fully defocused block.
Finally, the sum of the maximum difference in gradients
Because all the block-pairs are part of the source images, the (SMDG) can be calculated as Eq. (13).
gradient maps {gradi, i = 1, 2}, maximum gradient map gradmax XX
and minimum gradient map gradmin of the block-pair could be just SMDG ¼ ðgradmax ðx; yÞ  gradmin ðx; yÞÞ
copied from the corresponding part of {GMi, i = 1, 2}, GMmax and XX XX
¼ gradmax ðx; yÞ  gradmin ðx; yÞ: ð13Þ
GMmin, respectively. By this way, the gradients of the subregions
need not to be recomputed. Therefore, the proposed algorithm will In Eq. (13), SMDG of a block-pair is calculated as the sum of the
be accelerated. difference of the maximum gradient map gradmax and the mini-
Fourth, calculate the focus-measure of each block in the pair, as mum gradient map gradmin at each location. The expression of
the mathematical expression given below. the SMDG can be rewritten as the difference of the sum of the max-
XX imum gradients and the sum of the minimum gradients, as shown
FM i ¼ gradi ðx; yÞ; i ¼ 1; 2; ð9Þ
in Eq. (13). As mentioned previous, the maximum gradient map
where FMi is the focus-measure of the block i in the pair. gradmax approximates the gradient map of a fully focused block,
In Eq. (9), the gradient value gradi(x, y) is the focus-measure of and the minimum gradient map gradmin approximates the gradient
the pixel at the location (x, y) of the source image Ii. FMi simply map of a fully defocused block. Thus, SMDG of a block-pair approx-
sums up all the gradients in the block i to measure the focus of imates the difference in focus-measures of a fully focused block
the whole block i, which has been verified as an effective way in and a corresponding fully defocused block. According to the expla-
some commonly constructed focus-measures [3,9]. Though there nation of Eq. (12), if MDFM of a block-pair equals to the corre-
may exist some noise gradients in the gradient map gradi(x, y), sponding SMDG, then this block-pair fits Situation 1, in which a
the sum of the noise gradients in the block i would be very small, block is fully focused and another is fully out of focus. Also, this
compared with the corresponding focus-measure FMi. Thus, the relation of MDFM and SMDG is verified as followings.
focus-measure could partly suppress the influence of the noise According to Eqs. (6) and (9)–(12), MDFM of a block-pair, is actu-
gradients. ally equal to the absolute value of the sum of the difference of grad1
Fifth, find out the maximum and minimum focus-measures of and grad2 at each location. According to Eqs. (6)–(8) and (13), SMDG
the blocks in the pair, respectively. of the block-pair is equal to the sum of the difference of the greater
gradient of grad1 and grad2 and the smaller gradient of grad1 and
FM max ¼ maxfFM1 ; FM 2 g; ð10Þ grad2 at each location. Thus, MDFM of each block-pair is always less
than or equal to the corresponding SMDG. And only when the gra-
FM min ¼ minfFM 1 ; FM2 g; ð11Þ
dient grad1 is less than the gradient grad2 at each location or the
where FMmax and FMmin are the maximum and minimum focus- gradient grad1 is greater than the gradient grad2 at each location,
measures of all the block in the pair, respectively. MDFM of a block-pair will equal to the corresponding SMDG. That
In Eqs. (10) and (11), FMmax equals to the maximum focus- is, only when a block is fully focused and a block is fully defocused,
measure of the blocks in the pair, FMmin equals to the minimum MDFM of a block-pair will equal to the corresponding SMDG.
focus-measure of the blocks in the pair. For a partially focused block, Therefore, if MDFM of a block-pair equals to the corresponding
the focus-measure of this block will become greater, with the SMDG, there must be a fully focused block and a fully defocused
enlarging of the focused area in the block. And, the focus-measure block in the pair, i.e., this block-pair fits the Situation 1. After that,
of this block will become smaller, with the shrinking of the focused the focused block can be simply found out by comparing the focus-
area in the block. Thus, if a block-pair fits in Situation 1, in which a measures of the blocks in the pair. Otherwise, this block-pair fits
block is a fully focused and another is fully defocused, FMmax of this Situation 2, in which both blocks in the pair are partially focused.
block-pair would equal to the focus-measure of the fully focused Then, each block in this pair should be further subdivided in to four
block and FMmax should be the extreme maximum focus-measure quadrants.
of this block. Meanwhile, FMmin of this block-pair would equal to However in the real images, there may exist some noise pixels
the focus-measure of the fully defocused block and FMmin should in any one of the corresponding blocks. As a result, for a block-pair
be the extreme minimum focus-measure of this block. in Situation 1, MDFM of the block-pair usually appears a little less
Then, the maximum difference in focus-measures (MDFM) for than the corresponding SMDG. To suppress this defect, we choose
the block-pair can be calculated as the difference between 0.98  SMDG as the threshold value to discriminate the situation.
maximum focus-measure FMmax and minimum focus-measure For a block-pair, if MDFM P 0:98  SMDG, this block-pair fits Situ-
FMmin, as expressed in Eq. (12). ation 1. Otherwise, this block-pair fits Situation 2.
Hence, the proposed decomposition strategy is reasonable. In
MDFM ¼ FM max  FM min : ð12Þ
addition, in the proposed quadtree decomposition, the subdivision
In Eq. (12), FMmax represents the maximum focus-measure of of each block-pair only depends on the current block-pair itself
the corresponding blocks in the pair, and FMmin represents the rather than its parent block or the energy distribution of the image.
minimum focus-measure of the corresponding blocks in the pair. Therefore, the focused blocks could be detected more reliably by
The MDFM equals to the maximum difference in focus-measures the proposed strategy than De’s decomposition method.
X. Bai et al. / Information Fusion 22 (2015) 105–118 109

2.1.2. Detecting the focused blocks in a quadtree structure Fig. 5 shows a demonstration of the process of detecting the
In this part, we make use of the proposed quadtree decomposi- focused regions from the source images in a quadtree structure.
tion strategy to detect the focused blocks from m (m P 2) source In the detected focused regions (see (e and f)), the white regions
images. The detection process is described as follows. are the focused regions detected from the source image (a), the gray
At the beginning, the m source images are input as the root regions are the focused regions detected from the source image (b),
block-set at the first level of the quadtree structure. For the root and the black regions are the regions detected from both source
block-set, if MDFM P 0:98  SMDG, then this block-set fits Situa- images. It can be seen from (c) that the proposed algorithm could
tion 1 and the fully focused block can be found out by measuring well decompose the source images into blocks with optimal sizes
the focus of the blocks in this set. Otherwise, this block-set fits Sit- in a quadtree structure. And in this structure, the focused blocks
uation 2 and each block in the set should be further subdivided have been well detected by using the proposed algorithm, as shown
into four smaller blocks at the second level. This process is in (e). From (d), we can see that De’s decomposition method decom-
repeated on the smaller block-sets, until all focused blocks are poses the source images into almost all equal-sized blocks, rather
found out or the maximum level of the quadtree structure is than achieve the optimal subdivision of the source images. Similar
reached. Finally, the detected focused blocks are merged naturally to the block-based algorithms, the focused regions detected by De’s
to form the focused regions [3]. method yield the block artifacts, as shown in (f). In addition, the
In this paper, the source images are decomposed to the mini- source images are decomposed to a manual level by the De’s algo-
mum 2  2 sized blocks, and the corresponding level is the maxi- rithm, thus the performance of De’s algorithm would greatly
mum level, as calculated in Eq. (14). depends on the decomposition level, similar to the block-based
algorithms. Whereas, the proposed decomposition method decom-
max level ¼ logfmaxðM  1; N  1Þg þ 1; ð14Þ
pose the source images to the maximum level, and the subdivision
where M  N is the size of the source images; log is the logarithm is optimized. Thereby, the proposed quadtree-based detection
operator. method is more effective and robust than De’s method.
Before the quadtree decomposition, the dimensions of the
source images should be extended to a squared size and a power
2.2. A new weighted focus-measure: the sum of the weighted modified
of 2. In addition, the source images are input as the first level of
Laplacian
the quadtree structure. Therefore, if the minimum possible size
of the blocks is 2  2, the maximum level of the quadtree structure
In order to detect the focused regions precisely, a new weighted
equals to the integer part of log {max (M  1, N  1)} + 1.
focus-measure called the sum of the weighted modified Laplacian
When the source images are decomposed to the minimum
is proposed. The detailed information about the new focus-
2  2 sized blocks, it would be good enough for the precise detec-
measure is described as follows.
tion of the focused regions. Therefore, we choose its corresponding
For the multi-focus images, a focused block should preserve
decomposition level as the maximum level.
more sharp edges than the corresponding defocused blocks [3,9].
The process of detecting the focused blocks in a quadtree
In images, the gradients represent the edge information. The
structure is algorithmically summarized in Algorithm 1.
strength of the gradients could be used as a focus-measure, such
as the well-known focus-measures [9,14]: EOL, EOG, SML, etc.
Algorithm 1.
However in the homogeneous regions, the corresponding gradi-
ents are small. The focus-measures of both the focused block and
the corresponding defocused blocks would also be small. There-
(1) Read the m source images as the root block-set;
fore, the focus-measure of a focused block could not be clearly
(2) Compute the gradient map of each source image;
distinguished from the corresponding defocused blocks by using
(3) Calculate the maximum possible level (max level) of the
some commonly used focus-measures. And, the focus-measure of
quadtree structure;
a block in the homogenous regions could be easily affected by
(4) Decompose the source images into a quadtree structure:
the noise pixels.
Set current level l equal to 1;
To overcome this problem, we weight the gradient at each loca-
while (l < max level)
tion in a local window. By this way, the strength of the final gradi-
if the number of the block-sets at level l equals to zero
ent at each location would be enhanced by the nearby gradients.
then
For example, the gradients in the homogenous region would
break;
become larger, if there are sharp gradients around this region.
end
Therefore, the difference in focus-measures between a focused
for each block-set at level l
block and the corresponding defocused blocks would be increased.
Compute the corresponding SMDG and MDFM;
Then, the focus-measure of a focused block could be well distin-
if MDFM P 0.98  SMDG then
guished from the corresponding defocused blocks, either in the
Compute the focus-measure of each block in the set;
texture regions or in the homogeneous regions.
if only one block has the maximum focus-measure
In [9], SML has been experimentally proved to outperform the
then
other focus-measures. Hence, an assumption is made that the
Extract this block from the corresponding source
weighted version of SML would be effective. The experimental
image to the fusion image;
results prove that the weighted version of SML performs very well
end if
on measuring the focus. Therefore, we propose the weighted ver-
else
sion of SML, named as the sum of the weighted modified Laplacian
Subdivide each block in this set into four quadrants;
(SWML).
end if
next block-set
l = l + 1; 2.2.1. Sum of the modified Laplacian
loop In [14], Nayar et al. noted that: when calculating the Laplacian,
the second derivatives in the x and y directions may have the
110 X. Bai et al. / Information Fusion 22 (2015) 105–118

Fig. 5. A demonstration of the detection of the focused regions from the source images in a quadtree structure. (a–b) The source images; (c) the quadtree structure obtained
at the tenth level by the proposed quadtree decomposition strategy; (d) the quadtree structure obtained at the fifth level by De’s decomposition method; (e) the focused
blocks detected by the proposed algorithm; and (f) the focused blocks detected by De’s algorithm.

opposite signs, so that they tend to cancel each other. Therefore, a 2.2.2. Sum of the weighted modified Laplacian
small value may be obtained at a high gradient location. To sup- To improve the detection accuracy of the focused blocks in the
press this defect, Nayar et al. proposed the modified Laplacian homogenous regions, we present a weighted version of ML, named
(ML), which takes the absolute values of the second derivatives as the weighted modified Laplacian (WML).
in the Laplacian. The discrete approximation to ML can be First, the ML gradients of the source images are calculated
expressed as: according to Eq. (15). Then, the WML gradient at the pixel location
(x, y) is calculated as the weighted sum of the ML gradients in a
MLðx; yÞ ¼ j2f ðx; yÞ  f ðx  step; yÞ  f ðx þ step; yÞj
small local window. The WML can be mathematically expressed as
þ j2f ðx; yÞ  f ðx; y  stepÞ  f ðx; y þ stepÞj; ð15Þ
X
xþN X
yþN
MLði; jÞ
where step is a variable spacing, which is used to adapt to the pos- WMLðx; yÞ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; MLðx; yÞ > T;
2 2
i¼xN j¼yN 1 þ ði  xÞ þ ðj  yÞ
sible variations.
Then, the sum of the modified Laplacian (SML) at the location ð17Þ
(x, y) can be calculated in a small window around (x, y), as shown
in Eq. (16). where N determines the size of the window and T is a threshold
value.
X
xþN X
yþN
In Eq. (17), the WML gradient at the pixel location (x, y) is calcu-
SMLðx; yÞ ¼ MLði; jÞ; MLði; jÞ > T; ð16Þ
lated as the weighted sum of the ML gradients in a local window,
i¼xNj¼yN
the size of which is (2  N + 1)  (2  N + 1). In the WML gradient
where N determines the size of the window and T is a threshold map, the strength of the ML gradient at each location has been
value. propagated within a local window. With the increase of N, the size
X. Bai et al. / Information Fusion 22 (2015) 105–118 111

of the window will be increased and the propagation range of the Then, we only need to choose an appropriate value of N to gen-
ML gradient will be expanded. Also, the WML gradient at each erate the WML gradients of an image. From Eq. (17), we can know
location, which equals the weighted sum of all the ML gradients that the value of N determines the size of the weighted window,
in the local window, will become much greater. Especially for the which is (2  N + 1)  (2  N + 1). N also determines the propaga-
homogenous region, if there are sharp gradients around these tion range of the ML gradients in the WML gradient map. It is obvi-
regions within the local window, the gradients of the homogenous ous that the value of N should be affected by the size of the
regions would be strongly enhanced. Meanwhile, the related homogenous regions in the used image. If the image has
weight of the gradient at each location of the local window is large homogenous regions, like the background area in the ‘Lab’
inversely proportional to the spatial distance between the gradient image (see Fig. 1), the value of N should be large. If the image
location (i, j) and the center of the local window (x, y). Therefore, has small homogenous regions, such as the brick areas in the
the strength of the WML gradient at each location is mainly con- ‘Flower’ image (see Fig. 6(a and b)), the value of N should be small.
tributed by the nearby ML gradients. And the strength of the gradi- For a specified image, if the value of N is very small, the energy of
ents of the defocused region would not be influenced very much by the gradients could not be adequately propagated. If the value of N
the nearby focused regions. Overall, if an appropriate value of N is is very large, the gradients in the defocused region might be over-
chosen, then the WML gradients would be well distributed in the whelmed by the gradients in the focused regions. Different values
gradient map. Thereby it can be further used to accurately measure of N have been experimentally tested, ranging from 5 to 10. And we
the focus of the entire image. find that if the value of N equals to 8, the WML gradients would be
Then, the sum of the weighted modified Laplacian (SWML) at well propagated in most of the cases. Then, the new focus-measure
the location (x, y) can be calculated as: SWML could accurately measure the focus of the corresponding
blocks.
X
xþN X
yþN
SWMLðx; yÞ ¼ WMLði; jÞ; ð18Þ
i¼xNj¼yN 2.3. Reconstruction of the focused regions

where N determines the size of the window. Although the proposed quadtree detection method and the new
In Eq. (18), SWML of a window is computed as the sum of the weighted focus-measure WML are very effective, there are still
WML gradients in the window, similarly to SML. Here, we assume some small spurious regions inside the detected focused regions,
that an appropriate value of N has been chosen in Eq. (17), and then as shown in Figs. 5(e) and 6(h). It is obvious that the small isolated
the WML gradients of the source images could be properly distrib- blocks should not exist in a focused region [15]. The size of a
uted in the gradient maps. For the multi-focus images, the WML focused region must be much wider than the spurious blocks
gradients in a focused region could be much greater than the inside it [1,3]. In this paper, we make use of two sequential filters
WML gradients in the corresponding defocused regions, either in to reconstruct the focused regions. First, a morphological filter is
the texture regions or in the homogenous regions. Thus, SWML of used to process the focused regions. In this way, the small lines
a focused region could be clearly discriminated from the corre- or burrs could be eliminated and the nearby regions could be con-
sponding defocused regions. Also, because of the sum effect of nected. Then, a small region filter is applied to remove the small
SWML, this new focus-measure could partly suppress the noise isolated blocks in the focused regions detected from each source
gradients. Therefore, the new weighted focus-measure SWML image. Finally, the focused regions could be well extracted from
would be able to effectively measure the focus of the images. the source images and combined to form the final fusion image.
Fig. 6 shows a comparison example of SML and SWML. (a and b)
are the source images. (c and d) are the ML gradient maps of (a and 2.3.1. Morphological filter
b). (f and g) are the WML gradient maps of (a) and (b). From (c and In the detected focused regions, there may be some small lines
f), it can be seen that the strength of the ML gradients has been or burrs in the connection portions, and some nearby regions may
locally propagated in the WML gradient map. In the homogeneous be disconnected by the inappropriate decisions. Thus, an alternat-
regions, the strength of the gradients has been effectively ing opening and closing operators with a small disk structuring
enhanced, such as the flower area in (f). The same phenomenon element [1] is applied to process the focused regions detected from
can be seen from (d and g). In addition, we make use of SML and each source image. In this way, the small lines or burrs could be
SWML to detect the focused regions from the source images by eliminated, the connection portions of the focused regions could
the proposed algorithm. The detected regions are shown in (e and be smoothed, and the nearby regions would be combined as a
h), respectively. In the detected focused regions, the white regions whole region.
are the focused regions detected from the source image (a), the gray The morphological filter is used to do a minor adjustment of the
regions are the focused regions detected from source image (b), and focused regions. Thus, a small size of the disk structuring element
the black regions are the focused regions detected from both source would be enough for the deletion of the small burrs and the con-
images. Obviously, the detected focused regions by using SWML nection of the nearby regions. Also, different sizes of the disk struc-
(see (h)) have better consistency than SML (see (e)), especially in turing elements, with radius ranging from 2 to 5, have been used to
the homogeneous regions, such as the brick areas or the flower process the detected focused regions. We found that, when the
areas. It implies that the proposed focus-measure SWML performs radius of the disk structuring element equals to 3, the small lines
better than SML in the detection of the focused regions. or burrs could be well deleted and the nearby regions could be
right connected. Therefore, we use 3 as the radius of the disk
2.2.3. Parameter selection structuring element.
First, the settings of the parameters step and T in Eq. (15) are
given. ML is a commonly used gradient. In [9], the experiments 2.3.2. Small region filter
have verified the advantage of the SML over other focus-measures Usually, there may exist some small isolated regions inside a
in measuring the focus. In this paper, the corresponding parame- focused region. To handle this problem, we utilize a small region fil-
ters step and T are set equal to 1 and 5, as suggested in [9]. By this ter to delete the small isolated regions or fill the holes in the focused
way, the gradients introduced by the noise pixels could be regions detected from each source image [15]. Suppose that the size
effectively suppressed. And the ML gradients of the source images of the source images is M  N. Based on the visual observation, if the
could be well generated, such as the example shown in Fig. 6. area of an image region is less than (M  N)/40, this region is a very
112 X. Bai et al. / Information Fusion 22 (2015) 105–118

Fig. 6. A comparison example of SML and SWML. (a and b) The source images; (c and d) the ML gradient maps of the source images; (f and g) the WML gradient maps of the
source images; (e and h) the focus regions detected by using SML and SWML, respectively.

small region comparing with the whole image. Therefore, the small where Ii (i = 1, . . . , m) are the source images; GMi (i = 1, . . . , m) are
regions in this paper are defined as those regions whose area are the corresponding gradient maps of the source images.
less than (M  N)/40. The experimental results demonstrate that Because the proposed WML gradient is very effective and the
the small isolated regions could be effectively removed in most of corresponding gradient maps have been already achieved. Thus,
the cases by comparing the size of the region with (M  N)/40. After we adopt the proposed WML gradient here.
that, the consistency of the finally detected focused regions would Fig. 7 shows a demonstration example of the reconstruction
become better. process. (a and b) are the source images. (d) is the focused regions
detected by the proposed detection method. There are some iso-
2.3.3. Generating the fusion image lated blocks inside the detected focused regions. From (e), it can
The final fusion image can be produced as the following steps. be seen that the small lines are deleted and the nearby regions
are well connected by using the morphological filter. In (f), all
(1) The finally detected focused regions are copied directly from the small isolated regions are deleted by using the small region fil-
the source images to the fusion image. ter. Finally, the focused regions are well extracted and combined to
(2) There may exist some regions belonging to no image or more form one fully focused image, as shown in (c).
than one image, here the value of each pixel in these regions
is copied as the pixel from the source image i which has the 2.4. Comparison with the previous work
maximum gradient at this location. If the gradients of differ-
ent source images at a pixel location have the same value, Through the overall process of the proposed algorithm, it can be
then the value of the pixel at this location is copied as the concluded that compared with De’s algorithm, the proposed
average value of the pixels from all the source images. The algorithm mainly has two improvements: (1) presenting an effec-
mathematical expression is shown below [2]. tive quadtree decomposition strategy and (2) introducing a new
weighted focus-measure.
8 m
> X In De’s algorithm [3], the subdivision of each block-set greatly
>
> 1
Ii ðx; yÞ; if GMp ðx; yÞ ¼ GMq ðx; yÞ;
>m
> relies on the energy distribution of the source images and the cor-
>
>
< i¼1 responding parent block-set. Because the energy distribution of the
IF ðx; yÞ ¼ p – q; p; q 2 ½1; m ð19Þ image is always random, the optimal decomposition of the source
>
>
> I
> i ðx; yÞ; i ¼ arg maxðGMj ðx; yÞÞ; images usually could not be achieved. Whereas in the proposed
>
> j
>
: quadtree decomposition strategy, the subdivision of each block-
j 2 ½1; m
set only depends on the block-set itself, and the subdivision of each
X. Bai et al. / Information Fusion 22 (2015) 105–118 113

Fig. 7. A demonstration of the reconstruction process. (a and b) The multi-focus ‘Disk’ images; (c) the fusion image of the proposed algorithm; (d) the focused regions
detected by the proposed algorithm; (e and f) the focused regions processed sequentially by using the morphological filter and small region filter.

block-set is based on a reasonable approximation method. Thus, 3.1. Qualitative comparison


the source images could be optimally subdivided by the proposed
quadtree decomposition strategy. Therefore, the proposed quad- For the multi-focus image fusion, good visual effect means
tree decomposition strategy is more reasonable and reliable than effective fusion. To do the qualitative comparison, the performance
De’s decomposition method. Different from the general focus- of the algorithms should be evaluated by measuring the vision
measures, the proposed focus-measure uses the modified quality of the fusion images. Here, three comparison examples
Laplacian to suppress the noise interference and also uses the are given, as shown in Figs. 8–10 and some discussions on the
weighted method to properly propagate the energy of the gradient. experimental examples are stated as follows.
Thus, the focused blocks could be precisely detected by using the Fig. 8 shows the fusion results of the ‘Disk’ image set. It is obvi-
proposed focus-measure SWML, either in the texture regions or ous that the fusion images by the Gradient pyramid based
in the homogenous regions. Therefore, SWML is superior to the algorithm (Fig. 8(c)) and PCA based algorithm (Fig. 8(d)) appear
commonly used focus-measures, especially in the detection of low in contrast. The fusion images by the Gradient pyramid based
focused blocks in the homogenous regions. (Fig. 8(c)), PCA based (Fig. 8(d)), contrast pyramid based (Fig. 8(e)),
Moreover, different experiments have demonstrated that the DWT based (Fig. 8(f)) and SIDWT based (Fig. 8(g)) algorithms suffer
proposed algorithm is effective and outperforms some existing from the blurring effects, labeled by the yellow and orange rectan-
algorithms. gles. For De’s algorithm, its fusion image (Fig. 8(h)) yields blocking
artifacts and blurring edges, labeled by the yellow rectangles. The
3. Experimental results and discussions fusion image (Fig. 8(i)) produced by the proposed algorithm shows
the best visual effect.
To show the effectiveness of the proposed algorithm, we com- As is known, the contrast pyramid based, DWT based, gradient
pare our algorithm with some existing algorithms, including De’s contrast based, PCA based and SIDWT based algorithms are the
algorithm [3], contrast gradient based algorithm [16], Gradient transform domain techniques, and they would have a smooth
pyramid based algorithm [17], DWT based algorithm [18], PCA effect on the details of the image. Thereby, their fusion images
based algorithm [19], and shift invariant DWT based algorithm would suffer from the blurring effects around the edge area. As
(SIDWT) [20]. for De’s algorithm, their quadtree decomposition method may be
De’s algorithm is based on the quadtree structure. The Gradient not effective in some cases. In the experimental examples, the
pyramid, DWT, and SIDWT based algorithms are based on the source images are decomposed into almost all equal-sized blocks
multi-scale decomposition, so these algorithms are also decomposi- by De’s algorithm. Therefore, the fusion images of De’s algorithm
tion based algorithms. And, the PCA based algorithm is a commonly usually suffer from blocking artifacts. While the proposed algo-
used and effective algorithm for image fusion. Moreover, the pro- rithm could well extract the focused regions from the source
posed algorithm is an effective algorithm for multi-focus image images, and maintain them in good consistency, therefore the
fusion, through quadtree decomposition of the source images. fusion images would be produced with good visual quality. And
Thereby, using these algorithms as the comparison algorithms is the same conditions can be seen in Figs. 8 and 9.
reasonable. Fig. 9 shows the fusion results of the ‘Pepsi’ image set. The
In this paper, the experiments have been performed on several fusion images of the Gradient pyramid based algorithm (Fig. 9(c))
commonly used multi-focus image sets [1,3,8,9,12,13,26,27]. The and PCA based algorithm (Fig. 9(d)) appear to be reduced in con-
performance of the algorithms is evaluated by both the qualitative trast. The fusion images of the contrast pyramid based (Fig. 9(e)),
and quantitative methods. Some comparison examples are shown DWT based (Fig. 9(f)) and SIDWT based (Fig. 9(g)) algorithms have
below. the blurring effects around the text regions, labeled by the yellow
114 X. Bai et al. / Information Fusion 22 (2015) 105–118

Fig. 8. The multi-focus ‘Disk’ images and its fusion images. (a and b) The source images; (c–i) the fusion images by using contrast pyramid based algorithm, DWT based
algorithm, gradient pyramid based algorithm, PCA based algorithm, SIDWT based algorithm, De’s algorithm and the proposed algorithm, respectively.

rectangles. Also, the fusion image of De’s algorithm (Fig. 9(h)) has 3.2.1. Gradient similarity metric: QGM
the blocking artifacts in the character ‘P’ area, labeled by the yellow The first metric is the gradient similarity between the source
rectangle. In the end, the fusion image (Fig. 9(i)) of the proposed images and the fusion image. It is based on the observation that
algorithm is fully focused and good for the human visual a pixel has a high gradient value when it is focused [1,8]. For a
perception. set of multi-focus images {Ii, i = 1  m}, the gradient maps {GMi,
Fig. 10 shows the fusion results of the ‘Lab’ image set. The fusion i = 1  m} are firstly calculated. Then, these gradients are combined
images by the frequency domain algorithms (Figs. 10(c)–(g)) all into GM by taking the maximum gradient value at each location.
have the blurring effects around the head area, labeled by the yel- The mathematical expression can be described as
low rectangles. The fusion images of the gradient pyramid based
algorithm (Fig. 10(c)) and PCA based algorithm (Fig. 10(d)) appear GMðx; yÞ ¼ max fGMi ðx; yÞg: ð20Þ
i¼1;...;m
in low contrast. The fusion image (Fig. 10(h)) of De’s algorithm
Thereby, GM(x, y) has the highest gradient at each location. It
yields the blocking artifacts in some portions, labeled by the
could be used to approximate the gradient map of the original
orange rectangles. As expected, our algorithm produces a fusion
scene which is focused everywhere. Assuming that GM’ is the gra-
image with the best visual quality, as shown in Fig. 10(i).
dient map of the fusion image. Then, GM and GM0 are more similar,
The experimental examples show that the proposed algorithm
the performance of the corresponding algorithm is better. This gra-
performs better than the other algorithms. Also, it indicates that
dient similarity can be calculated as follows [1,3].
the proposed algorithm is competent for the fusion of the multi-
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
focus images. P 2
ðGMðx; yÞ  GM0 ðx; yÞÞ
Q GM ¼ 1  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
ffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi :
P ð21Þ
GM2 ðx; yÞ  GM02 ðx; yÞ
3.2. Quantitative comparison
As is known, the Robert gradient [21] is simple and easy for
To assess the fusion images objectively, the fusion algorithms computation. Because the Robert operator [21] performs very well
are also evaluated by three quantitative indices, which are gradient in the experiments, the Robert operator is used here.
similarity metric (QGM) [1,3], edge based similarity metric (QAB/F)
[21,22] and normalized mutual information (QMI) [23], respec- 3.2.2. Edge based similarity metric: QAB/F
tively. Their mathematical definitions are given below, followed An edge based similarity metric (QAB/F) proposed by Xydeas and
by the results and discussions. Petrovic [22,23], evaluates the amount of the edge information
X. Bai et al. / Information Fusion 22 (2015) 105–118 115

Fig. 9. The multi-focus ‘Pepsi’ images and its fusion images. (a and b) The source images; (c–i) the fusion images by using contrast pyramid based algorithm, DWT based
algorithm, gradient pyramid based algorithm, PCA based algorithm, SIDWT based algorithm, De’s algorithm and the proposed algorithm, respectively.

transferred from the source images to the fusion image. Generally, Usually, MI(X, Y) is normalized through dividing MI(X, Y) by
for two source images A and B, and a fusion image F, QAB/F is com- (H(X) + H(Y))/2 [25]. H(X) and H(Y) are the entropy of X and Y,
puted as respectively.
PN PM AF For the multi-focus image fusion, it has three variables: two
A
m¼1 ðQ ðn; mÞw ðn; mÞ þ Q BF ðn; mÞwB ðn; mÞÞ
Q AB=F ¼ n¼1
PN PM ð22Þ multi-focus images (IA and IB) and one fusion image (IF). QMI can
A þ wB ðn; mÞÞ
n¼1 m¼1 ðw ðn; mÞ be calculated as the sum of the normalized mutual information
between each source image and the fusion image. The mathemat-
where wA(n, m) and wB(n, m) are the local importance perceptual
ical expression is:
factors, QAF and QBF are the edge information preservation values.
In this paper, the parameters are used as suggested in [22]. MIðIA ; IF Þ MIðIB ; IF Þ
Q MI ðA; B; FÞ ¼ þ ð24Þ
ðHðIA Þ þ HðIF ÞÞ=2 ðHðIB Þ þ HðIF ÞÞ=2
3.2.3. Normalized mutual information: QMI
Mutual information (MI) is a quantitative measure of the
mutual dependence of two variables [24]. It shows the measure- 3.2.4. Quantitative comparison results
ment of the information shared by two images. Mathematically, The three quantitative indices are calculated according to their
MI between two discrete random X and Y variables is defined as definitions, and the results are shown in Tables 1–3. The bold
XX pðx; yÞ values are the maximum values in the corresponding rows. If the
MIðX; YÞ ¼ pðx; yÞ  log2 ð23Þ values of the three indices are larger, the performance of the
x2X y2Y
pðxÞ  pðyÞ
corresponding algorithm is better.
116 X. Bai et al. / Information Fusion 22 (2015) 105–118

Fig. 10. The multi-focus ‘Lab’ images and its fusion images. (a and b) The source images; (c–i) the fusion images by using contrast pyramid based algorithm, DWT based
algorithm, gradient pyramid based algorithm, PCA based algorithm, SIDWT based algorithm, De’s algorithm and the proposed algorithm, respectively.

Table 1
Gradient similarity metric.

Image sets Contrast pyramid DWT Gradient pyramid PCA SIDWT De’s algorithm Proposed algorithm
Clock 0.9735 0.9751 0.9621 0.9543 0.9819 0.9778 0.9762
Disk 0.9725 0.9755 0.9535 0.9239 0.9757 0.9725 0.9774
Flower 0.9648 0.9669 0.9445 0.9322 0.9645 0.9607 0.9671
Lab 0.9807 0.9818 0.9646 0.9463 0.9831 0.9813 0.9834
OpenGL 0.9556 0.9642 0.9376 0.8853 0.9662 0.9592 0.9608
Pepsi 0.9852 0.9854 0.9670 0.9449 0.9822 0.9860 0.9862

Table 2
Edge based similarity metric.

Image sets Contrast pyramid DWT Gradient pyramid PCA SIDWT De’s algorithm Proposed algorithm
Clock 0.6880 0.6619 0.6597 0.6137 0.7003 0.7384 0.7373
Disk 0.6849 0.6503 0.6440 0.5365 0.6822 0.7340 0.7383
Flower 0.6444 0.6222 0.6222 0.6197 0.6583 0.6821 0.6962
Lab 0.6910 0.6616 0.6585 0.5651 0.6839 0.7442 0.7470
OpenGL 0.6946 0.6784 0.6707 0.5893 0.7058 0.7289 0.7308
Pepsi 0.7548 0.7293 0.7342 0.6362 0.7445 0.7847 0.7847

The gradient similarity metric (QGM) measures the gradient sim- image. Compared with QGM, QAB/F could be more precise and
ilarity between the source images and fusion image. It can be seen reliable, because QAB/F is associated with both the strength and
from Table 1 that, the proposed algorithm performs the best in the orientation information of the edges. The comparison results in
most of the cases, which implies that the proposed algorithm can Table 2 show that the proposed algorithm performs better than
well preserve the sharp gradients from the source images. other algorithms in most of the cases. It indicates that the proposed
The edge based similarity metric (QAB/F) evaluates the amount of algorithm could well preserve the prominent texture information
edge information transferred from the source images to the fusion from the source images.
X. Bai et al. / Information Fusion 22 (2015) 105–118 117

Table 3
Normalized mutual information.

Image sets Contrast pyramid DWT Gradient pyramid PCA SIDWT De’s algorithm Proposed algorithm
Clock 0.9724 0.8827 0.8670 0.9703 0.9436 1.2075 1.2244
Disk 0.8275 0.7404 0.7466 0.8373 0.8033 1.1373 1.1459
Flower 0.7468 0.6659 0.6661 0.7566 0.7127 1.0171 1.1255
Lab 1.0015 0.9324 0.8953 1.0271 0.9935 1.2280 1.2456
OpenGL 0.9810 0.8914 0.7856 1.0279 0.9575 1.2714 1.2906
Pepsi 0.9643 0.8925 0.8656 0.9759 0.9289 1.2086 1.2483

Table 4
Comparison of the computational costs (s).

Image sets Contrast pyramid DWT Gradient pyramid PCA SIDWT De’s algorithm Proposed algorithm
Clock(500  503) 0.0523 0.0750 0.0800 0.0075 0.3072 1.5036 0.9946
Disk(640  480) 0.0613 0.0906 0.0973 0.0092 0.3922 2.3228 1.7222
Flower(764  576) 0.0846 0.1322 0.1412 0.0141 0.5952 3.1220 1.8370
Lab(640  480) 0.0630 0.0895 0.1018 0.0097 0.3961 2.2095 1.8393
OpenGL(758  569) 0.0895 0.1270 0.1424 0.0130 0.5713 2.7748 1.8204
Pepsi(512  512) 0.0520 0.0790 0.0836 0.0080 0.3208 1.5156 1.0290

Note: the size in each bracket represents the size of the corresponding source images.

The normalized mutual information (QMI) is a quantitative mea- detected from the source images in a quadtree structure, effec-
sure of the mutual dependence between the source images and the tively and precisely. And the detected focused regions could be
fusion image. It is clear from Table 3 that, the proposed algorithm well extracted from the source images and reconstructed to pro-
yields the best results in all the cases. This is because the proposed duce the fusion image.
algorithm can well combine the independent information from Because SWML is sensitive to the homogeneous regions, the
each source image to form one single image. Here, the independent focused regions could be detected more accurately by using SWML
information is referred as the focused regions detected from each than the commonly used focus-measures. In the proposed algo-
source image. rithm, all source images are decomposed to the maximum level
of the quadtree structure, therefore the focused regions could be
3.3. Comparison of the computational costs detected precisely in the tree structure. Moreover, the proposed
algorithm is simple, effective and reliable, because of the proposed
Finally, we have compared the computational costs of the quadtree decomposition strategy and the weighted focus-measure.
fusion algorithms. All the algorithms are running in a computer The experimental results show that the proposed algorithm per-
with Intel Processor (Intel Core i5-3470, 3.20 GHz) and 4 GB mem- forms very well for multi-focus image fusion.
ory. The computational times are listed in Table 4.
Table 4 shows that, the transform domain algorithms (the con- Acknowledgments
trast pyramid based, DWT based, gradient contrast based, PCA
based and SIDWT based algorithms) are all running fast. The spa- The authors thank the anonymous reviewers and editor very
tial domain methods (the proposed algorithm and De’s algorithm) much for their helpful and valuable comments, which substantially
are running slightly slower than the transform domain algorithms. improve the quality of the paper. The authors do acknowledge the
Also, it is obvious that the proposed algorithm is running faster source images used in their experiments. Multi-focus image sets
than De’s algorithm in all the cases. In the experimental examples, ‘Clock’, ‘Disk’, ‘Lab’ and ‘Pepsi’ are obtained from website [26].
the proposed algorithm can subdivide the source images optimally, Image sets ‘Flower’ and ‘OpenGL’ are obtained from website [27].
whereas De’s algorithm subdivides the source images into almost This work was partly supported by the National Natural Science
all equal-sized blocks. Thus, the decomposition times of the source Foundation of China (61271023), Program for New Century Excel-
images by using De’s algorithm would be a lot more than the pro- lent Talents in University (NCET-13-0020) and Fundamental
posed algorithm. And the computational cost of De’s algorithm will Research Funds for the Central Universities (YWF-14-YHXY-029,
be heavily increased with the increasing of its decomposition level. YWF-13-T-RSC-028, YWF-13-JQCJ-026).
Therefore, the proposed algorithm is faster than De’s algorithm.
Although the proposed algorithm has a little higher computational
cost than the transform domain algorithms, the proposed algo- References
rithm outperforms all the transform domain methods and the
[1] I. De, B. Chanda, B. Chattopadhyay, Enhancing effective depth-of-field by image
De’s algorithm from both the qualitative and quantitative evalua- fusion using mathematical morphology, Image Vis. Comput. 24 (12) (2006)
tions. Therefore, the proposed algorithm is competent for the 1278–1287.
multi-focus image fusion. [2] A. Saha, G. Bhatnagar, Q.M.J. Wu, Mutual spectral residual approach for
multifocus image fusion, Digit. Signal Process. 23 (4) (2013) 1121–1135.
[3] I. De, B. Chanda, Multi-focus image fusion using a morphology-based focus
measure in a quad-tree structure, Inf. Fusion 14 (2) (2013) 136–146.
4. Conclusion
[4] G. Qu, D. Zhang, P. Yan, Medical image fusion by wavelet transform modulus
maxima, Opt. Express 9 (4) (2001) 184–190.
This paper presents a simple yet effective quadtree-based algo- [5] X. Bai, F. Zhou, B. Xue, Fusion of infrared and visual images through region
rithm for multi-focus image fusion. In this work, we propose an extraction by using multi-scale center-surround top-hat transform, Opt.
Express 19 (9) (2011) 8444–8457.
effective quadtree decomposition strategy and also present a [6] F. Nencini, A. Garzelli, S. Baronti, L. Alparone, Remote sensing image fusion
new weighted focus-measure, thus the focused regions could be using the curvelet transform, Inf. Fusion 8 (2) (2007) 143–156.
118 X. Bai et al. / Information Fusion 22 (2015) 105–118

[7] A.A. Goshtasby, S. Nikolov, Image fusion: advances in the state of the art, Inf. [18] H. Li, B.S. Manjunath, S.K. Mitra, Multisensor image fusion using the wavelet
Fusion 8 (2) (2007) 114–118. transform, Graph. Model. Image Process. 57 (3) (1995) 235–245.
[8] S. Li, J.T. Kwok, Y. Wang, Combination of images with diverse focuses using the [19] P.S. Chavez, A.Y. Kwarteng, Extracting spectral contrast in Landsat thematic
spatial frequency, Inf. Fusion 26 (7) (2001) 169–176. mapper image data using selective principal component analysis,
[9] W. Huang, Z. Jing, Evaluation of focus measures in multi-focus image fusion, Photogramm. Eng. Remote Sens. 55 (3) (1989) 339–348.
Pattern Recogn. Lett. 28 (4) (2007) 493–500. [20] M. Unser, Texture classification and segmentation using wavelet frames, IEEE
[10] W. Huang, Z. Jing, Multi-focus image fusion using pulse coupled neural Trans. Image Process. 4 (9) (1995) 1549–1560.
network, Pattern Recogn. Lett. 28 (9) (2007) 1123–1132. [21] L.G. Roberts, Machine Perception of Three-Dimensional Solids, MIT Lincoln
[11] V. Aslantas, R. Kurbanan, Fusion of multi-focus images using differential Laboratory Technical Report, 1963.
evolution algorithm, Expert Syst. Appl. 37 (12) (2010) 8861–8870. [22] C. Xydeas, V. Petrovic, Objective image fusion performance measure, Electron.
[12] S. Li, B. Yang, Multifocus image fusion using region segmentation and spatial Lett. 36 (4) (2000) 308–309.
frequency, Image Vis. Comput. 26 (7) (2008) 971–979. [23] V. Petrovic, C. Xydeas, Sensor noise effects on signal-level image fusion
[13] L. Chen, J. Li, C.L.P. Chen, Regional multifocus image fusion using sparse performance, Inf. Fusion 4 (3) (2003) 167–183.
representation, Opt. Express 21 (4) (2013) 5182–5197. [24] A. Collignon, F. Maes, D. Vandermeulen, P. Suetens, G. Marchal, Automated
[14] S.K. Nayar, Y. Nakagawa, Shape from focus, IEEE Trans. Pattern Anal. Mach. multi-modality image registration based on information theory, Int. Conf.
Intell. 16 (8) (1994) 824–831. Inform. Process. Med. Imag. (1995) 263–274.
[15] Y. Zhang, L. Ge, Efficient fusion scheme for multi-focus images by using [25] T.O. Kvalseth, Entropy and correlation: some comments, IEEE Trans. Syst. Man
blurring measure, Digit. Signal Process. 19 (2) (2009) 186–193. Cybern. 17 (3) (1987) 517–519.
[16] A. Toet, L.J. Ruyven, J.M. Valeton, Merging thermal and visual images by a [26] http://www.quxiaobo.org/software/software_FusingImages.html
contrast pyramid, Opt. Eng. 28 (7) (1989) 789–792. [27] http://www.imgfsr.com/sitebuilder/images
[17] V.S. Petrovic, C.S. Xydeas, Gradient-based multiresolution image fusion, IEEE
Trans. Image Process. 13 (2) (2004) 228–237.

Potrebbero piacerti anche