3830 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO.

10, OCTOBER 2013

Image Size Invariant Visual Cryptography for
General Access Structures Subject to Display
Quality Constraints
Kai-Hui Lee and Pei-Ling Chiu
AbstractConventional visual cryptography (VC) suffers from
a pixel-expansion problem, or an uncontrollable display quality
problem for recovered images, and lacks a general approach
to construct visual secret sharing schemes for general access
structures. We propose a general and systematic approach to
address these issues without sophisticated codebook design. This
approach can be used for binary secret images in non-computer-
aided decryption environments. To avoid pixel expansion, we
design a set of column vectors to encrypt secret pixels rather
than using the conventional VC-based approach. We begin by
formulating a mathematic model for the VC construction prob-
lem to nd the column vectors for the optimal VC construction,
after which we develop a simulated-annealing-based algorithm
to solve the problem. The experimental results show that the
display quality of the recovered image is superior to that of
previous papers.
Index TermsVisual secret sharing scheme, pixel expan-
sion, general access structures, optimization, controllable display
quality.
I. INTRODUCTION
V
ISUAL cryptography (VC), which was proposed
by Naor and Shamir, allows the encryption of
secret information in image form [1]. Following their
work, much research was done on visual secret shar-
ing schemes (VSSs) [2]. From the point of view of
access structures, the existing VC schemes (VCSs) can
be divided into two categories: threshold access structure
(also known as k-out-of-n VCSs or (k, n)-VCSs) [3][5]
and general access structure (GAS) [6][12]. Naor and Shamir
focused on how to generate n shares such that the secret image
is revealed by at least k shares (2 k n).
Ateniese et al. (hereinafter Ateniese) [6] proposed the GAS
concept and also developed a VC-based solution for some
GASs. Using the GAS enables dealers to dene reasonable
combinations of shares as decryption conditions rather than
to specify the number of shares. For example, if there are
four participants (one president, one vice president, and two
Manuscript received July 26, 2012; revised March 2, 2013; accepted
April 29, 2013. Date of publication May 13, 2013; date of current version
August 28, 2013. This work was supported in part by the National Science
Council of Taiwan under Contracts NSC 101-2221-E-130-018 and NSC
101-2410-H-130-001. The associate editor coordinating the review of this
manuscript and approving it for publication was Prof. Jana Dittmann.
K.-H. Lee is with the Department of Computer Science and Informa-
tion Engineering, Ming Chuan University, Taoyuan 33348, Taiwan (e-mail:
khlee@mail.mcu.edu.tw).
P.-L. Chiu is with the Department of Risk Management and Insurance, Ming
Chuan University, Taipei 111, Taiwan (e-mail: plchiu@mail.mcu.edu.tw).
Digital Object Identier 10.1109/TIP.2013.2262290
managers) sharing a secret, the president may expect to decrypt
the secret with any single colleague who holds one of the
other shares, whereas the vice president is allowed to obtain
the secret only with two managers. The two managers are
restricted from accessing the secret. Given these exibilities,
we also can set the number of shares as the decrypting
condition. Clearly, (n, n)- and (k, n)-VCSs are special cases
of the GAS.
The pixel-expansion problem is a major drawback with
most VCSs that use the VC-based approach. The pixel-
expansion problem affects the practicability of a VC
scheme because it increases the storage and/or transmis-
sion costs. Moreover, the pixel-expansion problem usually
introduces the side effect that the recovered secret images
have less contrast. The contrast of the recovered images
decreases in proportion to 1/m, whereas the shares are
expanded by a factor of m times. As a result, the decrease
in contrast limits the application of these VC schemes.
To address the pixel expansion problem, Adhikari et al.
also proposed construction methods for VCSs for GASs;
their approach aims to reduce the pixel expansion factor for
(k, n)-VCSs [10]. Hsu et al. (hereinafter Hsu) used the proba-
bility concept to construct a VCS for GAS [7], [8]. However,
Hsus method does not guarantee enough blackness in some
access structures, such that recovered images cannot be recog-
nized by the naked eyes. Liu et al. proposed a deterministic
construction method for GASs to balance the drawbacks of
display quality and pixel expansion [9].
The display quality of a recovered image is affected not only
by its contrast value but also by its blackness [5], [13]. The
degree of blackness represents the probability that black secret
pixels will be accurately recovered. An image that has higher
contrast should have better display quality when the blackness
is xed [5]. However, a greater blackness value (e.g., 100%)
may decrease the contrast value of the recovered image. In
this paper, we take blackness as one of the design factors to
improve the display quality of recovered images.
In the past decade, a number of researchers have devel-
oped optimization models to address various VC construc-
tion problems [7], [8], [12], [13]. They have formulated the
threshold VCS for maximizing contrast or minimizing pixel
expansion as a linear programming problem. Koga proposed
a general formula to nd the basis matrices for (k, n)-VCSs
by the exhaustive search method; its objectives were to both
maximize the contrast and minimize the pixel expansion [13].
1057-7149 2013 IEEE
LEE AND CHIU: IMAGE SIZE INVARIANT VC FOR GASs 3831
Recently, Lee and Chiu (hereinafter Lee) proposed a generic
VC construction method for general access structures [12].
Their approach can perfectly recover black secret pixels, but
it results in decreasing contrast of recovered images in some
access structures. From our review, we note that the existing
VCS construction algorithms for GASs cannot simultaneously
avoid the pixel-expansion problem and guarantee an acceptable
blackness. These issues motivated us to develop a systematic
approach to the construction of size invariant VCSs (SIVCSs
or VCSs in short) for GASs subject to the adjustable display
quality of recovered images.
The proposed approach for SIVCSs is applicable to binary
secret images and no computational devices are needed during
the decryption phase. First, we formulate a mathematical
optimization model for the problem of the SIVCS for GASs
where the objectives are to maximize the worst and average
contrast of recovered images simultaneously under a blackness
constraint. Using this model, dealers can adjust the blackness
depending on the characteristics of the secret images to obtain
the best display quality for the recovered images. Then, we
develop a simulated-annealing-based algorithm to solve the
combinatorial optimization problem. Finally, we compare our
results with other approaches and present implementation
results to evaluate the effectiveness of the proposed algorithm.
The remainder of this paper is organized as follows:
Section II presents a review of background and related work.
In Section III, we introduce our model. The proposed opti-
mization model and the solution algorithm are introduced
in Section IV. In Section V, we show the results of an
experiment that was performed to evaluate the performance
of the proposed method. Finally, we summarize and conclude
our work in Section VI.
II. BACKGROUND AND RELATED WORK
A. Background of General Access Structures
Suppose P = {1, . . . ,n} is a set of n participants, and 2
P
denotes the power set of P. The quantity
Qual
denotes the set
of subsets of P from which we wish to share the secret; thus,

Qual
2
P
. Each set in
Qual
is said to be a qualied set,
and each set not in
Qual
is called a forbidden set (denoted
as
Forb
). Obviously,
Qual

Forb
= 2
P
and
Qual

Forb
=
. Based on these denitions, a VCS for an access structure
(
Qual
,
Forb
) on P can yield n shares. When we stack together
the shares associated with the participants in any set X

Qual
, we can recover the secret image, but any X
Forb
has no information on the stacked image.
The quantity
0
consists of all the minimal qualied sets:

0
= {A
Qual
: A

/
Qual
A

A}.
In traditional secret sharing schemes,
Qual
increases monoton-
ically and
Forb
decreases monotonically, the access structure
(
Qual
,
Forb
) is said to be strong, and
0
is called a basis. In
a strong access structure,

Qual
= {C P: B C for some B
0
},
and we say that
Qual
is the closure of
0
. If
Qual
=
0
,
then the access structure (
Qual
,
Forb
) is said to be weak.
For example, if three participants share a secret image (i.e.,
P = {1, 2, 3}) and
0
= {{1, 2} , {1, 3}}, in the strong access
structure (
Qual
,
Forb
)-VCS, stacking any set of subsets {1, 2},
{1, 3}, or {1, 2, 3} can reveal the secret image; otherwise, no
information can be displayed. However, in the weak access
structure (
0
,
Forb
)-VCS, only stacking sets {1, 2} and {1, 3}
can reveal the secret image; the image is not guaranteed to be
revealed with set {1, 2, 3}.
B. Review of VCSs for GASs
1) Atenieses Approach: In 1996, Ateniese rst proposed
a VC-based approach for VCSs for GASs. They mapped a
VCS access structure to a graph and found both the lower and
upper bounds on the size of the shares (i.e., the pixel-expansion
factor) from the graph. They gave minimum pixel-expansion
factors as well as basis matrices for VCSs for strong access
structures for a maximum of four participants [6].
MacPherson extended Atenieses model to include grey-
scale images and derive new results on the minimum possible
pixel expansion for all possible GASs on at most four par-
ticipants. However, a method for constructing the grey-scale
VCSs for GASs remains an open problem [11].
As with other conventional VC-based approaches, the
above-mentioned VCS approach for GASs also suffers from
the pixel-expansion problem. There also are other drawbacks
with Atenieses approach. For example, black secret pixels
cannot be completely recovered, the aspect ratio of the recov-
ered image cannot be maintained, and this approach needs a
sophisticated codebook design.
2) Hsus Approach: In 2006, Hsu, for the rst time, reported
the formulation of an unexpanded VCS for a GAS problem
as an optimization model [7], [8]. Their method adopts a set
of n 1 column vectors to share a secret pixel to encrypt n
participants, thus eliminating the drawbacks of pixel expan-
sion. Based on the model, a probability matrix can be found
and used to encrypt a secret for a specic access structure.
Hsus objective is to maximize contrast values for all qualied
recovered images subject to the security constraint. They use
the goal-programming technique and also develop a genetic-
based algorithm to solve the optimization problem [7], [8].
Hsus approaches have better maximum and average con-
trast values for recovered images than Atenieses results in
some cases. However, Hsus method still has problems. First,
the approach is probabilistic, which leads to poor visual quality
for the recovered secret images when the blackness of the
images is low. Second, Hsus objective cannot guarantee an
acceptable contrast level for recovered images in the worst
case [8].
3) Lees Approach: Lee proposed the formulation of a
SIVCS for strong general access structures, (
Qual
,
Forb
),
based on the probabilistic (n, n)VCSs [12]. Lees approach
is to nd the quantity of basis shares and a construction set for
a given access structure. The basis shares that were yielded by
the probabilistic (n, n)VCSs are used to synthesize the shares
of (
Qual
,
Forb
)-VCS according to the construction set.
For example, suppose there are four participants, P =
{1, 2, 3, 4}, who share a secret image and the qualied set
3832 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 10, OCTOBER 2013
is
Qual
= {{1, 2} , {1, 2, 3} , {1, 2, 4} , {1, 3, 4} , {1, 2, 3, 4}}.
The VCS for (
Qual
,
Forb
) can be constructed by 3 basis
shares and construction set C = {{s
1
} , {s
2
, s
3
} , {s
2
} , {s
3
}}.
Hence, the encryption procedure generates 3 basis shares
s
1
, s
2
, and s
3
by utilizing the construction of an existing
(3, 3)-VCS. Then, by construction set C, the procedure con-
structs 4 shares, S
1
, S
2
, S
3
, and S
4
, for the (
Qual
,
Forb
)-VCS.
That is, assign basis shares s
1
, s
2
, and s
3
to shares S
1
, S
3
,
and S
4
, respectively; Stack shares s
2
and s
3
together and then
assign the stacked share to S
2
. Obviously, upon stacking all
shares in any qualied set in
Qual
, the secret image can be
restored as same as the (3, 3)-VCS can do.
Lees approach utilizes pixel-expansion-free (n, n)-VCS
to synthesize the strong (
Qual
,
Forb
)-VCS, therefore, the
approach also can avoid the pixel expansion problem. Lees
approach has the following drawbacks. First, it relies on
existing (n, n)-VCS, so the visual quality of recovered images
depends on the (n, n)-VCS. Second, the encryption process
needs two phases to generate shares.
III. THE PROPOSED MODEL
The main idea behind the proposed SIVCS is the probabilis-
tic visual cryptography (ProbVC) which was rst proposed by
Ito et al. in 1999 [4]. Ito et al. constructed the (k, n)-VCS
by using two collections of column vectors, C
0
and C
1
, which
are transformed from basis matrices of the conventional (k, n)-
VCS. Suppose the basis matrix contains nm entries, C
1
(C
0
)
will contain m n 1 column vectors. To share a black (white)
pixel, one of the column vectors in C
1
(C
0
) is randomly chosen
and then distributes i -th entry in the column vector to i -th
share. For example, the (2, 3)-ProbVC scheme is constructed
by the following collections of column vectors
C
0
=

1
1
1

0
0
0

0
0
0

and
C
1
=

1
0
0

0
1
0

0
0
1

C
1
and C
0
are transformed from two 33 basis matrices of the
conventional (2, 3)-VCS. For encrypting a black secret pixel
and suppose the chosen column vector is
_
0 1 0
_
T
, the pixels
0, 1, and 0 are distributed to shares 1, 2, and 3, respectively.
In this fashion, each secret pixel within a secret image is
encrypted in only one pixel in each constituent share. Thus,
image size of shared and stacked images is same as the secret
image.
The approach of Ito et al. has to rely on existing basis
matrices of the conventional VCSs. To relax the limitation,
Yang proposed general construction rules for (k, n)-ProbVCS
in 2004 [3]. Both of Ito et al. and Yang have proved the
ProbVC is as secure as the conventional VCSs.
In the present study, we develop a general construction
methodology for SIVCSs for GASs. To state our concept
formally, we rst present our denitions as follows.
Denition 1: The n-tuple Boolean column vector S = [s
j
]
T
with 1 j n, is dened as an encoding pattern for
each original pixel, where s
j
= 0(1) denotes that the pixel is
encoded as a white (black) sharing pixel in share j .
Denition 2: Suppose P = {1, . . . ,n} is a set of n partic-
ipants and 2
P
denotes the power set of P. When a set of
participants X with X 2
P
\, stack their shares (which were
encrypted by vector S) together, the visual effect (i.e., black or
white) of a stacked pixel can be obtained by L(v
X
) = s
p
1
+
s
p
2
+. . .+s
p
k
, where k = |X| and p
1
,, p
k
X. The quantity
v
X
is a k-tuple column vector, v
T
X
= [s
p
1
s
p
2
. . .s
p
k
], and the
operator + represents the OR operation. If L(v
X
) = 1 (0),
the corresponding pixel will be decoded as black (white) on
the stacked image.
Example 1: Suppose P = {1, 2, 3} and two participants
stack their shares together. At a coordinate, pixel values on
their shares are black, white, and white, respectively. The
column vector of participants 1 and 2 is v
T
{1,2}
=
_
s
1
s
2
_
=
_
1 0
_
and their stacked pixel is black, which can be calculated
by L
_
v
{1,2}
_
= s
1
+s
2
= 1. The column vector of participants
2 and 3 is v
T
{2,3}
=
_
s
2
s
3
_
=
_
0 0
_
and their stacked pixel is
white, which can be calculated by L
_
v
{2,3}
_
= s
2
+s
3
= 0.
Blackness is an important factor for recovered images.
The blackness of a binary image denotes the appearance fre-
quency of black pixels in the image. In VC-based approaches,
the blackness of a recovered image is proportional to the
Hamming weight H(V) of the OR-ed m-vector V, which
is stacked by m shared subpixels of participants [1]. In a
conventional (2, 2)-VCS, the basis matrices for white and
black secret pixels are
_
0 1
0 1
_
and
_
0 1
1 0
_
; hence, the blackness
of white and black secret pixels are H
__
0 1
__
= 1 and
H
__
1 1
__
= 2, respectively. Similarly, in this study, the
blackness of recovered image can be determined by the OR-
ed operation of all column vectors [3].
Denition 3: Suppose P = {1, . . . ,n} is a set of n par-
ticipants that share an image encoded by an encoding set
C = {S
i
, 1 i m}, where S
i
= [s
i, j
] is an n-tuple Boolean
column vector. For any subset set X,
_
i
1
, i
2
, . . . , i
q
_
is all
members of X with
_
i
1
, i
2
, . . . , i
q
_
{1, 2, . . . , n}. Let V
C,X
denotes the collections of q-tuple vectors that are obtained by
restricting each n-tuple vector in C to rows i
1
, i
2
, . . . , i
q
. The
set
C,X
=
_
L
_
v
i,X
_
, v
i,X
V
C,X
, X 2
P
\
_
represents
the stacked result of all shares in X. The blackness of
this recovered image is H
_

C,X
_
/m, where H
_

C,X
_
is the
Hamming weight of
C,X
.
The contrast for the recovered image is one of the major
metrics for evaluating the performance of VCSs. For a recov-
ered (binary) image of VCSs, the contrast of the image is the
difference in the blackness of recovered pixels for black and
white secret pixels. Contrast can be dened as below.
Denition 4: Suppose an image is stacked from shares held
by a set of participants X. These shares hold a portion of a
secret image via a VC scheme and two collections C
0
and C
1
of sets, where C
t
= {S
i
, 1 i m}, t {0, 1}. The contrast
(denoted as
X
) and blackness (denoted as
X
) for the image
can be dened as

X
=
H
_

C
1
,X
_
H
_

C
0
,X
_
m
and
X
=
H
_

C
1
,X
_
m
.
LEE AND CHIU: IMAGE SIZE INVARIANT VC FOR GASs 3833
Denition 5: Suppose P = {1, . . . ,n} is a set of n par-
ticipants. A solution with non-expandable shares to the VCS
for an access structure (
Qual
,
Forb
) on P consists of two
collections of sets, C
0
and C
1
of sets, where C
t
= {S
i
, 1
i m}, t {0, 1}. Collection C
0
(C
1
) is used to encrypt
white (black) pixels of secret images. Suppose
TH
(
TH
>0)
is the threshold for detecting a difference in an image by the
humans visual system. The solution is considered feasible if
the following conditions are satised:
1) For any Y
Forb
, V
C
1
,Y
= V
C
0
,Y
.
2) For any X
Qual
, H
_

C
1
,X
_
H
_

C
0
,X
_
>
TH
.
Condition 1 is the security condition on restricting secret
accessibility of any forbidden set. For any forbidden set
Y, assume
_
i
1
, i
2
, . . . , i
q
_
is all members of Y and V
C
0
,Y
(V
C
1
,Y
) denotes the collection of column vectors that are
obtained by restricting each n-tuple vectors in C
0
(C
1
) to
rows i
1
, i
2
, . . . , i
q
. Then V
C
0
,Y
and V
C
1
,Y
have to contain
the same collections of q-tuple vectors with the same chosen
probabilities (i.e., H
_

C
1
,Y
_
/m = H
_

C
0
,Y
_
/m). The black
and white secret pixels are therefore indistinguishable by
humans visual system. The property is equivalent to the secu-
rity condition presented by Naor and Shamir for conventional
VCSs [1]. Hence, condition 1 is valid as a security condition
of ProbVCSs.
Condition 2 ensures that the blackness of recovered black
secret pixels is higher than that of recovered white secret pixels
in a qualied recovered image. If H
_

C
1
,Y
_
H
_

C
0
,Y
_
>
TH
,
a humans visual system can recognize a difference between
the recovered secret pixels. If
TH
is large enough, a humans
visual system can distinguish between the recovered black and
white secret pixels to obtain the secret images.
Using these denitions, a SIVC scheme for access structure
(
Qual
,
Forb
) can be constructed as follows: let two collections
of sets C
0
and C
1
be adopted for the SIVCS. In the encryption
phase, the dealer randomly chooses one column vector from
C
0
(C
1
) to encrypt white (black) secret pixels.
The above-mentioned method, which is also called the
single pixel encoding method encrypts a secret image pixel by
pixel. This method is easy and low in complexity, but it cannot
guarantee that the pixel can be uniformly distributed in a small
area in the reconstructed image. It may decrease the quality
of reconstructed images [17][18]. Chow et al. proposed a
multi-pixel block size invariant VCS that maintains the relative
pixel density in a small area in the reconstructed image to
improve the quality of the image [17]. The proposed scheme
is built from existing basis matrices of the conventional VCS.
However, in Chow et al.s scheme, the encryption process is
performed by taking a multi-pixel block as a unit of encryp-
tion. They suggest that to use the pixel expansion factor, m,
in the conventional VCS as the block size of encryption.
Zhang et al. also worked on the multi-pixel encoding method
to improve the quality of the reconstructed image [18]. Their
method is similar to Chow et al.s; however, it collected the
pixel block in the secret image by a zigzag scan method in
each encoding run. We focus on how to nd code collection
sets, C
0
and C
1
, for SIVCSs upon a given access structure.
Hence, we use the single pixel encoding method to generate
shares in the following experiments.
Example 2: Suppose there are four participants P =
{1, 2, 3, 4}, that share a secret image; the minimal qualied
set is
0
= {{1, 2, 3} , {1, 4} , {3, 4}} and the forbidden set
is
Forb
= 2
P
\
Qual
. Then, the (
Qual
,
Forb
)-VCS can be
constructed using the following two sets of collections
C
0
= {2 :E
0
, 1 : E
6
, 1 : E
11
, 1 : E
13
}
=

0
0
0
0

0
0
0
0

0
1
1
0

1
0
1
1

1
1
0
1

and
C
1
= {1 : E
1
, 1 : E
2
, 1 : E
5
, 1 : E
8
, 1 : E
14
}
=

0
0
0
1

0
0
1
0

0
1
0
1

1
0
0
0

1
1
1
0

.
We rst examine the security condition for forbidden sets.
Let Y = {1, 2}
Forb
, then by Denition 3, we have

C
0
,Y
=
_
L
_
v
1,Y
_
, L
_
v
2,Y
_
, L
_
v
3,Y
_
, L
_
v
4,Y
_
, L
_
v
5,Y
__
=
_
L
__
0
0
__
, L
__
0
0
__
, L
__
0
1
__
,
L
__
1
0
__
, L
__
1
1
___
= {0, 0, 1, 1, 1} ,

C
1
,Y
=
_
L
__
0
0
__
, L
__
0
0
__
, L
__
0
1
__
,
L
__
1
0
__
, L
__
1
1
___
= {0, 0, 1, 1, 1} ,
and
V
C
1
,Y
= V
C
0
,Y
=
__
0
0
_
,
_
0
0
_
,
_
0
1
_
,
_
1
0
_
,
_
1
1
__
.
Hence, the forbidden set Y = {1, 2} is secure. It is easy to
verify that all forbidden sets meet the security condition.
Next, we calculate the contrast condition for qualied set
X = {1, 2, 3}. We have
C
0
,X
= {0, 0, 1, 1, 1} and
C
1
,X
=
{0, 1, 1, 1, 1}. Thus, the contrast of the recovered image, which
is stacked from shares 1, 2, and 3, is

X
=
H
_

C
1
,X
_
H
_

C
0
,X
_
m
=
4 3
5
=
1
5
.
In the same way, we have contrasts
{1,4}
= 2/5 and
{3,4}
=
1/5. The contrast of the recovered image that is stacked from
the other qualied sets also can have
X
1/5, X
Qual
.
This shows that collections C
0
and C
1
are valid solutions
for the (
Qual
,
Forb
)-VCS because it meets the security and
contrast conditions.
In Example 2, notation E
T
j
=
_
e
j ,1
, . . . ,e
j,n
_
, 0 j < 2
n
,
denotes a n-tuple Boolean column vectors for encrypting n
shares, where

n
i=1
2
ni
e
j,i
= j . Notation q : E
T
j
, q 1,
q Z, denotes q column vectors E
T
j
. In the rest of this paper,
3834 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 10, OCTOBER 2013
TABLE I
THE USED CODE COLLECTIONS IN EXAMPLE 4
Code Collections
C
0
C
1
Set p {2 :E
0
, 1 : E
7
, 1 : E
9
, 1 : E
14
} {1 : E
1
, 1 : E
2
, 1 : E
4
, 1 : E
8
, 1 : E
15
}
Set q {1 : E
0
, 2 : E
7
, 1 : E
9
, 1 : E
14
} {1 : E
3
, 1 : E
5
, 1 : E
6
, 1 : E
8
, 1 : E
15
}
Set r {1 : E
0
, 2 : E
7
, 1 : E
9
, 2 : E
14
} {1 : E
3
, 1 : E
5
, 1 : E
6
, 1 : E
10
, 1 : E
12
, 1 : E
15
}
Fig. 1. Relationship between
0
,
Qual
, and
Forb
.
we use the notation for denoting the collection of column
vector. By Example 2, we make the following observation.
Observation 1: In a (
Qual
,
Forb
)-VCS, the contrast of
each recovered image can vary, and depends on the access
structure of each qualied set.
In the previous studies on visual cryptography, researchers
tried to construct a (
Qual
,
Forb
)-VCS such that the secrets
can be revealed while all shares in qualied set X, X
Qual
,
were stacked. The relationship between
0
,
Qual
, and
Forb
is as shown in Fig. 1. Qualied set
Qual
is the closure of
0
,
that is, any qualied set X, X
Qual
and X /
0
, should
be a subset of Y, Y
0
. That means the secret also can be
revealed by stacking all shares in minimal set Y. On the other
hand, the quantity of
Qual
is no less than the quantity of
0
.
Hence, to nd the encodings that satisfy the contrast condition
for all sets in
0
could be easier than that for all sets in
Qual
.
Denition 6: Suppose P = {1, . . . ,n} is a set of n partici-
pants. A solution of a (
0
,
Forb
)-VCS on P consists of two
collections of sets C
0
and C
1
of sets. Collection C
0
(C
1
) is
used to encrypt white (black) pixels of secret images. The
solution is considered feasible if the security condition can be
satised for any set Y
Forb
and the secret can be revealed
for any set X
1

0
. The contrast condition can be ignored
for the qualied set X
2
, X
2

Qual
and X
2
/
0
. The (
0
,

Forb
)-VCS also is called a weak VCS for GASs.
Example 3: Suppose P = {1, 2, 3, 4} and the minimal
qualied set is
0
= {{1, 2, 3} , {1, 4} , {3, 4}}. Qualied set

Forb
is the closure of
0
and the forbidden set is
Forb
=
2
P
\
Qual
. Then, the (
0
,
Forb
)-VCS can be constructed using
two sets of collections:C
0
= {1 : E
0
, 1 : E
6
, 1 : E
11
, 1 : E
13
}
and C
1
= {1 : E
2
, 1 : E
5
, 1 : E
9
, 1 : E
14
}.
By Denition 6, it is easy to determine if the code collec-
tions can meet the security condition; that is, V
C
1
,Y
= V
C
0
,Y
,
Y
Forb
. The contrast of the recovered image that is stacked
from qualied set X, X
0
, is
X
= 1/4.
The (
0
,
Forb
)-VCS ignores the contrast condition for
qualied set X, X
Qual
and X /
0
; therefore, stacking
all shares in X cannot reveal the secret. For example, in
TABLE II
THE CONTRAST AND BLACKNESS OF RECOVERED IMAGES IN EXAMPLE 4
Qualied
Sets X
Contrast (
X
) Blackness (
X
)
Set p Set q Set r Set p Set q Set r
{1, 2, 4} 1/5 1/5 1/6 4/5 1 1
{1, 3, 4} 1/5 1/5 1/6 4/5 1 1
{2, 3} 1/5 1/5 1/6 3/5 4/5 1
Example 3, the recovered image by stacking shares 1, 2, and
4 cannot reveal the secret because the image cannot meet
the contrast condition (i.e.,
{1,2,4}
= 0, {1, 2, 4}
Qual
and
{1, 2, 4} /
0
). However, it has an alternative way to reveal
the secret by stacking shares 1 and 4 (i.e.,
{1,4}
= 1/4,
{1, 4}
0
).
From Example 3, we have the following second observation.
Observation 2: Instead of applying the (
Qual
,
Forb
)-VCS,
the (
0
,
Forb
)-VCS can have a higher opportunity to construct
recovered images with better contrast.
Example 4: Suppose there are four participants P =
{1, 2, 3, 4}, that share a secret image; the minimal qualied
set is
0
= {{1, 2, 4} , {1, 3, 4} , {2, 3}} and the forbidden set
is
Forb
= 2
P
\
Qual
. Dealers use three code collections, as
listed in Table I, to construct the (
0
,
Forb
)-VCS. The contrast
and blackness values of the recovered image of each minimal
qualied set are listed in Table II.
In this example, code sets p and q have same contrast
values for all minimal qualied sets; however, all blackness
of code set q are better than those of code set p. That
means that the recovered images constructed by code set q
have a better display quality than those constructed by code
set p. Although dealers would like to further improve the
blackness value of the recovered images, doing so may have
the undesirable side effect of decreasing the contrast value
of the image. For example, dealers would like to construct
the (
0
,
Forb
)-VCS subject to the 100% blackness for all
recovered images by using code set r; however, this results
in the contrast decreasing from 1/5 to 1/6 for all images, by
comparison with code set q.
Observation 3: Blackness of a recovered image will affect
(i.e., increase or decrease) the display quality of the image.
The above-mentioned observations indicate two major
issues in VCS construction for GASs. First, the contrast of
the images in a qualied set may vary across a wide range,
which, in the worst case, will lead to poor display quality.
Second, the display quality of the recovered images can be
LEE AND CHIU: IMAGE SIZE INVARIANT VC FOR GASs 3835
improved by adjusting the blackness of the recovered images.
We address these problems in the following section.
IV. PROPOSED ALGORITHM
A. Formulation
The problem for constructing a (
0
,
Forb
)-VCS is to nd
the two collections of sets C
0
and C
1
subject to the security,
contrast, and blackness constraints. From the perspective of the
quality of the recovered image, the objectives of the proposed
formulation are to maximize the worst and the average contrast
for recovered secret images among all minimal qualied sets.
The proposed problem can be formulated as a multi-
objective optimization model given below.
Given parameters:
P Set of participants, P = {1,. . ., n}.
n Number of participants, n 2.
Control parameter for blackness of recovered images,
Z.

0
Set of minimal qualied sets,
0
2
P\
.

Qual
Qualied set,
Qual
closure(
0
).

Forb
Set of forbidden sets,
Forb
= 2
P
\
Qual
.
Decision variables:
m Number of column vectors in code collection C
0
(or C
1
),
m 2.
S
i
n-tuple Boolean column vector, S
i
= [s
i, j
], 1 i m,
1 j n; s
i, j
= 0 (1) indicates that the pixel is encoded
as a white (black) sharing pixel in share j .
C
0
C
0
= {S
i
, 1 i m}, consists of a set of column
vectors used to share a white pixel.
C
1
C
1
= {S
i
, 1 i m}, consists of a set of column
vectors used to share a black pixel.

X
Contrast for a recovered image that is reconstructed
by shares held on a set of participants X, X 2
P
\;
contrast
X
can be dened as
X
=
H
_

C
1
,X
_
H
_

C
0
,X
_
m
.
Objective function:
maximize mi n
X
0

X
maximize
1
|
0
|

X
0

X
(IP1)
subject to:
V
C
1
,Y
= V
C
0
,Y
, Y
Forb
(1)
H
_

C
1
,Y
_
> 0, Y
Forb
, |Y| = 1 (2)
H
_

C
1
,X
_
H
_

C
0
,X
_
> 0, X
0
(3)
m H
_

C
1
,X
_
, X
0
(4)
s
i, j
{0, 1}, 1 i m, 1 j n (5)
m 2, m Z. (6)
The rst objective of the formulation is to maximize the
contrast of the worst recovered image in the minimal qualied
set. The second objective is to promote the average contrast of
all recovered images in the minimal qualied set. Constraint
(1) ensures that the solution for (
0
,
Forb
)-VCS obeys the
security condition. Constraint (2) species that each share
must contain some black pixels. Constraint (3) guarantees that
each qualied recovered image can have a nonzero contrast
such that the image can be recognized. Constraint (4) ensures
that the minimal blackness for all black secret pixels can be
reconstructed. In other words, the blackness of all recovered
images must be at least equal to (m )/m at least. By
adjusting constant , dealers can achieve better display quality
for recovered images. Constraints (5) and (6) limit the ranges
of the decision variables.
The complexity of the solution space for the optimization
problem is extremely high. The column vector S
i
has 2
m
possible combinations to construct a white or black secret
pixel for n shares. Therefore, the complexity to solve the
optimization problem is O(2
2mn
).
B. Solution Approach
In this section, we proposed an algorithm to solve
the proposed multi-objective optimization problem for the
(
0
,
Forb
)-VCS. To simplify solution procedures and improve
the efciency of the proposed algorithm, some constraints of
the original optimization model are relaxed before the solution
procedures start. Then, the proposed simulated-annealing-
based (SA-based) algorithm is developed to solve the opti-
mization problem.
1) Model Transformation: First, we relax Constraints (1),
(3), and (4) of the proposed model by penalizing the objective
function. The penalty function P for code collections C
0
and
C
1
, can be dened as follows:
P (C
0
, C
1
) = 1 +

Y
Forb
S(Y)
+

|
0
|

X
0
(C(X) D(X))

where
S(Y) =
_
1+S
F
(Y) , V
C
1
,Y
= V
C
0
,Y
0, otherwise
C(X) =
_
1, H
_

C
1
,X
_
H
_

C
0
,X
_
>0
0, otherwise
and
D(X) =
__
m H
_

C
1
,X
__
/m, m H
_

C
1
,X
_
>
0, otherwise.
Function S(Y) determines the number of forbidden sets
that violates the security constraint. Parameter , =
0.5 |
0
|/|
Forb
|, which is the weight of function S(Y),
is used to balance the penalty between the qualied sets
and the forbidden sets. The fraction part of S(Y), S
F
(Y) =
_
H
_

C
1
,Y
_
H
_

C
0
,Y
__
/m, which is an external penalty
function, is used to indicate the degree of the security con-
straint violation for a code collection. Functions C(X) and
D(X) calculate the number of minimal qualied sets that
meets the contrast condition (i.e., Constraint (3)) and the
blackness constraint (i.e., Constraint (4)). Function D(X) is
used to measure the difference between the blackness of a
qualied recovered image and its target value. If the image
cannot meet the blackness constraint, the difference will be a
penalty for the image. When Constraints (1), (3), and (4) were
3836 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 10, OCTOBER 2013
satised, the function values of S(Y), C(X), and D(X) should
be 0, |
0
|, and 0, respectively. Hence, function P for a feasible
solution must be equal to1.
Next, the multiple objective functions can be combined to
simplify the complexity of the solution procedure. In this
study, we take the rst objective in IP1 as the primary goal.
The combined objective function F for code collection sets C
0
and C
1
is dened as follows:
F(C
0
, C
1
) =
_
P (C
0
, C
1
) , P (C
0
, C
1
) > 1

/K, P (C
0
, C
1
) = 1
where

= K
_ _
K mi n
X
0

X
_
+ (1/|
0
|)

X
0

X
_
and K is a big constant (e.g., 10000). The rst
equation of

, mi n
X
0

X
, which is equivalent to the rst
objective in IP1 represents the worst contrast of recovered
image. The value of
_
K mi n
X
0

X
_
is greater than 1 if
the code collection is feasible and constant K is large enough.
The second equation of

,
1
|
0
|

X
0

X
, which equals to
the second objective in IP1 is less than 1. Hence, the rst
objective becomes the integer part of function F and the second
objective is the fraction part of function F. Obviously, function
F will be less than 1 if the code collection is feasible.
Eventually, the original multi-objective optimization model
can be transformed as IP2:
Objective function:
minimize F(C
0
, C
1
) (IP2)
subject to:
H
_

C
1
,Y
_
>0, Y
Forb
, |Y| = 1 (7)
s
i, j
{0, 1}, 1 i m, 1 j n (8)
m 2, m Z. (9)
Original model IP1 can be replaced by the penalized model
(IP2). In this way, Constraints (1), (3), and (4) can be omitted
during the solution process. Constraints (7), (8), and (9) are
the same as Constraints (2), (5), and (6), respectively.
2) The Proposed Algorithm: The SA approach is a generic
and probabilistic meta-heuristic method for solving difcult
optimization problems. It can solve combinatorial optimization
problems in a large search space; that is, it can nd good
approximated solutions to the global optimums by randomized
heuristic methods. In this section, we develop a SA-based
algorithm to solve the proposed mathematical optimization
formulation (IP2) for the GAS problem.
The relationship between the pixel expansion and the con-
trast in the conventional VC-based VCSs is sophisticated and
remains an open question. Yang proved that the required
column vectors in the probabilistic VCSs are equivalent to
the pixel expansion factor in conventional VCSs [3]. Hence,
to determine the required number of column vectors m in the
proposed model IP2 is a difcult problem. Koga exhaustively
searched for solutions for threshold VCSs within a given range
of m and found the best one for the VCSs [3]; this approach
is inefcient. An efcient method for nding a minimal m
within a given range is proposed in the following.
Lemma 1: Suppose code collection C = {C
0
, C
1
} is the
optimal solution of a VCS and no redundant column-vectors
TABLE III
PSEUDO CODE OF PROPOSED REDUCE_CV ALGORITHM
Algorithm Reduce_CV()
Input: C, m
Output: C, m
1. 1 i m, r
0,i
0, r
1,i
0
2. 1 i m, if S
0
i
=S
1
j
and r
1, j
= 1 (S
0
i
C
0
, S
1
j
C
1
, 1 j m)
then r
0,i
1, r
1, j
1
3. C C, m m
4. 1 i m, if r
0,i
= 1 then delete S
0
i
from C
0
i
5. 1 i m, if r
1,i
= 1 then delete S
1
i
from C
1
i
, m m 1
6. Output C, m
exist in C. Column vectors S C
0
and S C
1
are called
redundant vectors in C.
Proof: Suppose C
1
and C
0
consisting m column vectors,
is an optimal solution of the VCS, column vector s C
0
and S C
1
. Obviously, vector s cannot alter the security and
value of H
X
to the VCS where H
X
= H
C
1
,X
H
C
0
,X
and X
is a qualied set. Therefore, vector s can be deleted from C
0
and C
1
simultaneously. The contrast value of set X will be
altered from H
X
/m to H
X
/(m 1). In other words, the new
code collection in which redundant vector s was deleted has
a higher contrast value for set X than the original one. That
is contradictory.
Theorem 1: A feasible code collection C

can be reduced to
an optimal code collection C by deleting all redundant column
vectors from C

.
Proof: Based on Lemma 1, the theorem holds obviously.
Theorem 1 indicates that an optimal solution of a VCS can
be found when the given range of m is larger than its minimal
value.
Example 5: Suppose there are four participants P =
{1, 2, 3, 4},
0
= {{1, 2, 3} , {1, 4}}, the optimal values of the
(
0
,
Forb
)-VCS are m = 4, the optimal contrasts
{1,2,3}
=
1/4, and
{1,4}
= 1/2. A feasible code collection C

=
_
C

0
, C

1
_
, C

0
= {1 : E
0
, 1 : E
6
, 1 : E
11
, 1 : E
13
, 1 : E
3
, 1 :
E
10
} and C

1
= {1 : E
3
, 1 : E
5
, 1 : E
8
, 1 : E
14
, 1 : E
3
, 1 : E
10
},
can be obtained when we nd the solution in the case of m= 6.
Hence, we have the contrasts
{1,2,3}
= 1/6 and
{1,4}
=
1/3. By Theorem 1, the right-most two column vectors can be
deleted. Therefore, we can obtain the optimal contrast values
1/4 and 1/2 for sets {1, 2, 3} and {1, 4}, respectively.
Based on Theorem 1, the value of m is no longer a
sensitive parameter in the VCS construction. As illustrated in
Example 5, given a moderate value of m, a feasible solution
for a specic VC scheme can be found, and then the best
construction for the VCS can be obtained by reducing the
column vectors of the feasible solution.
The pseudo code for deleting redundant column vectors in a
feasible solution is listed in Table III. Step 1 resets all markers
r
0,i
(r
1,i
), 1 i m, that are used to indicate whether or not a
column vector i is redundant in C
0
(C
1
). Step 2 marks a pair of
redundant column vectors in C
0
and C
1
as redundant vectors
by setting corresponding markers. Steps 4 and 5 delete all
redundant vectors in C according to the markers. Finally, the
LEE AND CHIU: IMAGE SIZE INVARIANT VC FOR GASs 3837
TABLE IV
PSEUDO CODE OF PROPOSED SA-BASED ALGORITHM
Algorithm GAS_SA()
Input: n, m,
0
,
Forb
Output: C
best
, f
1. 1 i n, randomly generate an initial guess for code collection
C = {C
0
, C
1
} such that H
_

C
0
,{i}
_
= H
_

C
1
,{i}
_
= m/2
2. Calculate energy E for the above initial guess
3. E
old
E, E
best
E
old
, C
best
C
4. t t
0
, r r
0
5. While t t
f
do
6. Repeat r times
7. Randomly select a share i and a column j , let s
0
i, j
1 s
0
i, j
,
s
0
i, j
C
0
8. Randomly select a column j

, where s
1
i, j

= 1 s
0
i, j
and
s
1
i, j

C
1
, let s
1
i, j

s
0
i, j
9. Calculate E for the new conguration
10. If E < 1 then
11. Call Reduce_CV (C, m, C, m)
12. Y
Forb
, |Y| = 1, if H
_

C
1
,Y
_
= 0 then goto Step 7
13. Calculate E based on m
14. EndIf
15. E
new
E
16. E E
new
E
old
17. Generate a random number uniformly distributed in [0, 1).
18. If E < 0 or < e
(E/t )
then
19. Z
old
Z
new
20. If E
old
< E
best
then E
best
E
old
, C
best
C
21. else recover the action in Steps 7 and 8
22. End Repeat
23. t
T
t, r
T
r
24. End While
25. If E
best
1 then f 0, goto Step 29
26. f 1
27. Call Reduce_CV (C
best
, m, C, m)
28. C
best
C
29. Output C
best
, f
output is a reduced code collection C and number of column
vectors in C
0
(i.e., m).
The SA-based algorithm, as listed in Table IV, is developed
for solving model IP2. The proposed approach treats decision
variable m as a given variable and tries to nd the best
solution with a given access structure (
0
,
Forb
) based on
m. Step 1 randomly guesses an initial value for C
0
and
C
1
subject to the security constraint on each share. Step 2
calculates energy value E for the rst solution. Given that the
optimization problem in IP2 is a minimization problem, the
energy function in the SA-based algorithm is directly dened
as E = F(C
0
, C
1
). Steps 3 and 4 initialize related parameters
for the SA procedure.
Steps 524 are the main SA loop, which will be terminated
when the frozen temperature, t
f
, is reached. Steps 622 are
executed r times to rene solutions in a state of equilibrium.
Steps 7 and 8 randomly explore a next solution by altering
the current solution state of its neighborhood. The step size
in Step 7 is very small, the algorithm selects column vector j
in C
0
, and then alters the encoding of share i . The altered
encoding is denoted by s
0
i, j
. Step 7 indicates that share i
violates the security condition, which is corrected in Step 8.
Step 8 selects a column vector in C
1
in which the encoding
of share i , s
1
i, j

, differs from s
0
i, j
and then alter s
1
i, j

. In such
a way, the encoding of share i can be altered and the security
condition of share i can be preserved. Steps 1014 deal with
Constraint (7) in model IP2 only while the current solution
receives no penalty (i.e., E < 1). Step 11 reduces the current
code collection and then Step 12 checks whether or not the
code collection meets Constraint (7). If the constraint can be
satised, the energy of the feasible solution is reevaluated
based on the actual quantity of vectors in the solution (i.e., m)
in Step 13; if the constraint cannot be satised, then it aborts
the solution. Steps 1521 evaluate the value of the energy
function for the new state and decide whether or not the current
state will be replaced by the new state. The objective value
with the minimum energy value should be saved as the best
solution in Step 20. After r solution iterations, parameters t
and r are modied in Step 23.
The solution procedure is terminated when t < t
f
. When the
algorithm stops, if E
best
< 1, the best-found solution, C
best
, is
a feasible solution to this problem. Steps 2628 reduce C
best
and set a feasibility indicator of C
best
as feasible (i.e., f = 1).
In contrast, if E
best
1, it will reset indicator f in Step 25.
Finally, the algorithm outputs indicator f and the best-found
solution C
best
as the solution to the problem.
The running time of the proposed SA-based algorithm
contains two parts: outer and inner loops. The outer loop
contains steps from Step 5 to Step 24. The running of the
outer loop depends on the parameter setting for the cooling
scheduler of SA process. The inner loop contains steps from
Step 7 to Step 21. Step 9 dominates the time complexity of
the loop. The calculation of energy function E involves all
decryption possibilities of n participants. The time complexity
to check the security condition for a forbidden set is O(m).
Hence, the time complexity of the inner loop is O(m2
n
).
V. EXPERIMENTAL RESULTS
A. Scenario
In the following subsections, we assess the performance
of the proposed algorithm from the quantitative and the
qualitative viewpoints. From the quantitative viewpoint, we
examine the performance of the proposed scheme and solution
approach. In this study, we use the worst and the average
contrasts of recovered images as quantitative performance
metrics. The worst contrast (
min
) and the average contrast
(
avg
) of a minimal qualied set
0
are dened as:

min
= mi n
X
0

X
and
avg
=
1
|
0
|

X
0

X
.
From the qualitative viewpoint, we evaluate our results by
visual effects. In both evaluations, we test access structures
as listed in Table V. The parameters for the algorithm are
t
0
= 0.05, t
f
= t
0
/15,
T
= 0.75, and
T
= 1.1. The other
parameters for the algorithm also are listed in Table V. We
also compare our results with other approaches.
B. (
0
,
Forb
)-VCS vs. (
Qual
,
Forb
)-VCS
First, we assess the performance of (
0
,
Forb
)-
VCSs and (
Qual
,
Forb
)-VCSs in terms of the worst
contrast
min
, the average contrast
avg
, and the
minimal required number of column vectors m. As
shown in Table VI, the
min
of (
0
,
Forb
)-VCSs
3838 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 10, OCTOBER 2013
TABLE V
SOME MINIMAL QUALIFIED SETS AND ITS RELATED PARAMETERS USED
IN THIS EXPERIMENT
No n
0
Parameters for SA
r
0
m
1 4 2-out-of-4 2, 000 20
2 4 123,14 100 20
3 4 123,14,34 100 20
4 4 134,12,23,24 100 20
5 4 123,124 200 20
6 4 124,134,23 200 20
7 4 123,124,134 200 20
8 4 3-out-of-4 200 20
9 5 1234,1235,1245,345 20, 000 40
10 5 1234,1345,235 20, 000 40
11 5 1235,1245,234,345 10
5
40
12 5 4-out-of-5 10
5
40
13 6 1235,126,1345,1356,2345 310
5
40
14 6 1245,1246,1346,12356,3456 510
5
40
15 7 146,12357,34 1.210
6
40
16 7 1234,1457,2567,3467 1.510
6
40
17 7 1234,147,2567,3467 10
6
40
18 7 1234,34567 210
6
40
19 8 13478,23,256 10
6
40
20 8 12,1357,178,236,2468 10
6
40
is better than that of (
Qual
,
Forb
)-VCSs in some minimal
qualied sets, which are denoted by bold font. This indicates
that a (
0
,
Forb
)-VCS can produce recovered images with
a higher contrast value than a (
Qual
,
Forb
)-VCS. The

avg
of
_

Qual
,
Forb
_
-VCSs is slightly larger than that of
(
0
,
Forb
)-VCSs in access structure 3. However, the minimal
required m for a (
0
,
Forb
)-VCS can be smaller than a
(
Qual
,
Forb
)-VCS needs. That implies a (
0
,
Forb
)-VCS has
a smaller solution space than a (
Qual
,
Forb
)-VCS. Hence,
we evaluate the performance of the proposed algorithm only
for (
0
,
Forb
)-VCSs in the following experiments.
C. Performances of (
0
,
Forb
)-VCSs
In this subsection, we investigate how the blackness
constraint (i.e., Constraint (4)) affects performances of a
(
0
,
Forb
)-VCS. The evaluation is performed under three
optimization models:
Model A: (
0
,
Forb
)-VCS without blackness constraint;
that is, Constraint (4) is omitted.
Model B: (
0
,
Forb
)-VCSs under the blackness con-
straint with blackness control parameter = 1.
Model C: (
0
,
Forb
)-VCSs under the blackness con-
straint with parameter = 0.
The worst-case contrasts of these models are listed in
Table VII. From Table VII, we make the following observa-
tions:
1. In optimization Model A, recovered images can reach
100% blackness for all minimal qualied sets in some
access structures; for example, access structures 4 and 5.
TABLE VI
A COMPARISON BETWEEN (
0
,
Forb
)- AND (
Qual
,
Forb
)-VCSS
No
(
0
,
Forb
)-VCS (
Qual
,
Forb
)-VCS
m
min

avg
m
min

avg
1 6 1/3 33.3% 6 1/3 33.3%
2 4 1/4 37.5% 4 1/4 37.5%
3 4 1/4 25.0% 5 1/5 26.7%
4 4 1/4 25.0% 4 1/4 25.0%
5 4 1/4 25.0% 4 1/4 25.0%
6 5 1/5 20.0% 5 1/5 20.0%
7 6 1/6 16.7% 6 1/6 16.7%
8 6 1/6 16.7% 6 1/6 16.7%
9 12 1/12 10.4% 12 1/12 10.4%
10 10 1/10 10.0% 10 1/10 10.0%
11 10 1/10 10.0% 10 1/10 10.0%
12 15 1/15 6.7% 15 1/15 6.7%
13 14 1/14 7.1% 15 1/15 6.7%
14 18 1/18 5.6% 18 1/18 5.6%
15 18 1/18 18.5% 19 1/19 12.3%
16 18 1/18 5.6% 25 1/25 6.0%
17 16 1/16 6.3% 21 1/21 6.0%
18 20 1/20 5.0% 24 1/24 4.2%
19 19 1/19 19.3% 19 1/19 5.3%
20 14 1/14 17.2% 15 1/15 12.0%
m: number of column vectors in code collection C
0
(or C
1
)
In the general cases, recovered images cannot be guaran-
teed to reach 100% blackness for all minimal qualied
sets in most of the access structures. In access structures
3, 6, 10, 11, 15, 17, and 19, the blackness of the
worst-case recovered images was no more than 80%.
The worst-case recovered images means the image
has the worst contrast among an access structure. The
low blackness value affects the display quality of a
recovered image; in fact, the recovered image may not
be recognizable when the contrast of the image is too
low.
2. In optimization Models B and C, the blackness of the
worst-case recovered image can be improved in some
access structures without decreasing the contrast value
of the image. For example, in access structures 2, 7,
and 1520, the blackness value of all images can be
promoted to 100% (i.e., = 0) without losing the
contrast values of the images. Except in the case of
access structure 9, the blackness value of all images
can be promoted to (m 1)/m (i.e., = 1) without
decreasing the contrast values of the images. In access
structure 19, the blackness can be increased up to 36%
by Model B.
3. In the majority of the cases (access structures 1, 3, 6,
and 814 in Table VII), the contrast values (
min
)
will decrease under optimization Model C to construct
(
0
,
Forb
)-VCSs. The decrease cannot lead to a serious
problem in recognizing the contents of the recovered
images, but the increased m value will enlarge the search
space of the optimization problem.
LEE AND CHIU: IMAGE SIZE INVARIANT VC FOR GASs 3839
TABLE VII
THE WORST-CASE CONTRAST (
min
) OF (
0
,
Forb
)-VCSS IN VARIOUS
BLACKNESS CONSTRAINTS
No
Optimization Models
A B ( = 1) C ( = 0)
m
min
m
min
m
min
1 6 1/3 5/6 6 1/3 8 1/4
2 4 1/4 1 4 1/4 4 1/4
3 4 1/4 3/4 4 1/4 9 2/9
4 4 1/4 1 4 1/4 4 1/4
5 4 1/4 1 4 1/4 4 1/4
6 5 1/5 3/5 5 1/5 6 1/6
7 6 1/6 5/6 6 1/6 6 1/6
8 6 1/6 5/6 6 1/6 9 1/9
9 12 1/12 5/6 15 1/15 18 1/18
10 10 1/10 4/5 10 1/10 12 1/12
11 10 1/10 4/5 10 1/10 14 1/14
12 15 1/15 13/15 15 1/15 25 1/25
13 14 1/14 6/7 14 1/14 18 1/18
14 18 1/18 8/9 18 1/18 25 1/25
15 18 1/18 7/9 18 1/18 18 1/18
16 18 1/18 5/6 18 1/18 18 1/18
17 16 1/16 3/4 16 1/16 16 1/16
18 20 1/20 17/20 20 1/20 20 1/20
19 19 1/19 11/19 19 1/19 19 1/19
20 14 1/14 6/7 14 1/14 14 1/14
: the blackness of the recovered image which has the worst contrast
value
min
Table VIII lists the average contrast (
avg
) of the
(
0
,
Forb
)-VCSs in Table VI. As listed in Table VIII, except
for access structure 9, Models A and B have the same average
contrast for all access structures, which indicates that the
blackness constraint ( = 1) cannot affect the average contrast
of recovered images. When the model tries to achieve the
highest blackness (i.e., Model C, = 0) for all recovered
images, the average contrast decreases slightly in some access
structures (e.g., access structures 1, 3, 8, 9, and 1114).
However, in the case of access structure 6, Model C has the
highest average contrast and blackness among all models at
the same time. This case shows that the blackness constraint
(especially, in the case of = 0) can affect (increase or
decrease) the average contrast of the images only slightly.
In summary, the blackness constraint can increase the dis-
play quality of the recovered images, but it also can decrease
the display quality and increase the number of required column
vectors, which rapidly increases the solution space. The effect
of the constraint is complicated and highly depends on the
access structure of a VCS. Hence, combining the blackness
constraint into the VCS optimization model can help the dealer
in nding a code collection for producing recovered images
that have better display quality by adjusting the blackness
control parameter .
D. Comparison With Other Approaches
Next, we compare our results (Model A) with the results of
Ateniese [6], Hsu [7], and Lee [12]. Hsu reported his results
TABLE VIII
THE AVERAGE CONTRAST (
avg
) OF (
0
,
Forb
)-VCSS IN VARIOUS
BLACKNESS CONSTRAINTS
No
Without
Blackness
Constraint
With
Blackness Constraint
= 1 = 0
m
avg
m
avg
m
avg
1 6 33.3% 6 33.3% 8 25.0%
2 4 37.5% 4 37.5% 4 37.5%
3 4 25.0% 4 25.0% 9 22.2%
4 4 25.0% 4 25.0% 4 25.0%
5 4 25.0% 4 25.0% 4 25.0%
6 5 20.0% 5 20.0% 6 22.2%
7 6 16.7% 6 16.7% 6 16.7%
8 6 16.7% 6 16.7% 9 11.1%
9 12 10.4% 15 10.0% 18 8.3%
10 10 10.0% 10 10.0% 12 10.0%
11 10 10.0% 10 10.0% 14 7.1%
12 15 6.7% 15 6.7% 25 4.0%
13 14 7.1% 14 7.1% 18 6.7%
14 18 5.6% 18 5.6% 25 4.0%
15 18 18.5% 18 18.5% 18 18.5%
16 18 5.6% 18 5.6% 18 5.6%
17 16 6.3% 16 6.3% 16 6.3%
18 20 5.0% 20 5.0% 20 5.0%
19 19 19.3% 19 19.3% 19 19.3%
20 14 17.2% 14 17.2% 14 17.2%
only for the access structures 27; hence, our comparison
focuses on these access structures.
Fig. 2 shows that the proposed approach can achieve better
contrast values in the worst case of recovered images than
Hsus probabilistic constructions (i.e., access structures 3,
4, 6, and 7), Atenieses VC-based constructions (i.e., access
structure 3), and Lees approach (i.e., access structures 3, 4, 6,
and 7). These results prove the effectiveness of the proposed
optimization model in improving the contrast for the image in
the worst case.
From Fig. 2 and 3, the proposed approach has the same
values of
min
and
avg
with Atenieses results except those
for access structure 3. It indicates that both approaches have
found the best results for access structures 2 and 47. From
the results of Hsus and this study in access structures 3 and 4,
these access structures have conicting objectives between
maximizing the worst and average contrasts. Hsus aimed
to maximize the average contrast, therefore his approach
decreases the contrast value in the worst case in access
structures 3 and 4. Except for the above-mentioned cases,
the results of this study have better performances than other
approaches in terms of
min
and
avg
.
In access structure 3 (access structure is
0
=
{{1, 2, 3}, {1, 4}, {3, 4}}), qualied set {1, 2, 3} has the worst
recovered image than other sets. Hence, we use the recov-
ered image of set {1, 2, 3} as an example in the following
experiment. Fig. 4 shows a comparison between the proposed
study and other approaches by visual effects. The recovered
images in Fig. 4(c) and (e) are difcult to recognize due to the
3840 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 10, OCTOBER 2013
Fig. 2. Comparison by
min
.
Fig. 3. Comparison by
avg
.
(a) (b) (c) (d)
(e)
Fig. 4. Comparison between other approaches and the proposed study on the
worst case result (i.e., set {1, 2, 3}) of access structure {{1, 2, 3}, {1, 4}, {3, 4}},
(a) Secret image with 96 64 pixels (192 DPI), (b) the recovered images of
this study (contrast
min
= 2/9, blackness = 1), (c) the recovered image
of Hsus study (
min
= 0.15, = 0.75), (d) the recovered image of Lees
study (
min
= 1/8, = 1), (e) the recovered image of Atenieses study (pixel
expansion factor = 5,
min
= 1/5, = 0.8).
low contrast as well as blackness. As shown in Fig. 4(b), the
proposed approach can produce the clearest recovered image
for set {1, 2, 3} than other approaches. The code collection for
producing Fig. 4(b) is listed in Table IX (access structure 3,
= 0).
There are two major differences between Hsus model
and the proposed models. First, the proposed model aims to
maximize contrast for recovered images in the worst case.
In contrast, Hsus model focuses on promoting the average
contrast for recovered images. Second, the proposed model
guarantees the minimal blackness for recovered images, but
Hsus model does not. These differences indicate that the pro-
posed model can produce better display quality for recovered
images in the worst case.
A part of the solutions for the four participants produced
by the proposed algorithm are listed in Table IX. These
solutions can help readers to verify the correctness of the
following comparison results. More results are available on
http://www.csie.mcu.edu.tw/khlee/vc/gas.htm.
TABLE IX
A PART OF VCS SOLUTIONS FOR ACCESS STRUCTURES LISTED IN
TABLE V
No m Code Collects
2 4 0
C
0
= {1 : E
0
, 1 : E
6
, 1 : E
11
, 1 : E
13
}
C
1
= {1 : E
3
, 1 : E
5
, 1 : E
8
, 1 : E
14
}
3 9 0
C
0
=
{1 : E
0
, 1 : E
1
, 1 : E
6
, 1 : E
7
, 3 : E
11
, 1 : E
12
, 1 : E
13
}
C
1
= {2 :E
3
, 2 : E
5
, 2 : E
9
, 1 : E
10
, 2 : E
14
}
3 4 1
C
0
= {1 : E
0
, 1 : E
6
, 1 : E
11
, 1 : E
13
}
C
1
= {1 : E
2
, 1 : E
5
, 1 : E
9
, 1 : E
14
}
4 4 0
C
0
= {1 : E
0
, 1 : E
7
, 1 : E
13
, 1 : E
14
}
C
1
= {1 : E
5
, 1 : E
6
, 1 : E
11
, 1 : E
12
}
5 4 0
C
0
= {1 : E
0
, 1 : E
7
, 1 : E
11
, 1 : E
12
}
C
1
= {1 : E
3
, 1 : E
4
, 1 : E
8
, 1 : E
15
}
6 5 1
C
0
= {1 : E
0
, 2 :E
7
, 1 : E
9
, 1 : E
14
}
C
1
= {1 : E
3
, 1 : E
5
, 1 : E
6
, 1 : E
8
, 1 : E
15
}
7 6 0
C
0
= {1 : E
0
, 2 : E
7
, 1 : E
11
, 1 : E
13
, 1 : E
14
}
C
1
= {1 : E
3
, 1 : E
5
, 1 : E
6
, 1 : E
8
, 2 : E
15
}
: Control parameter for blackness of recovered images.
E. Demonstrations and Discussions
In this subsection, we show the implementation results
of access structure 10, which has ve participants
and shares a secret image based on access structure
{{1, 2, 3, 4}, {1, 3, 4, 5}, {2, 3, 5}}. The cipher-text in the
secret image used in this experiment, as shown in Fig. 5(a),
was written in various font sizes (i.e., 220, 160, 84, 72,
60, 48, 36, and 24 pts). Fig. 5(b), (c), and (d) demonstrate
the worst-case recovered images of Models A, B, and C,
respectively.
Fig. 5(b) shows that the worst-case recovered image of
Model A can achieve only an 80% blackness; thus, the cipher-
texts in smaller font sizes (e.g., the upper part of Fig. 5(b))
look very dim and are very difcult for the human eye to
recognize. The situation indicates that a VC scheme without
the blackness constraint may fail to decrypt secrets from the
recovered image if the cipher-texts are not in a large/bold font
and that the recovered black secret pixels cannot remain at a
high blackness. This case demonstrates that a VCS may fail
to decrypt secret images for some access structures due to the
poor visual quality of the worst-case recovered images, even
if the contrast of the image is maximized.
Fig. 5(c) and (d) illustrate that the cipher-texts become clear
when the blackness is increased. In Fig. 5(c), four S in a
larger font can be recognized. In Fig. 5(d), all S are clear
enough for decrypting except for the smallest one. Although
the contrast value of Fig. 5(d) is lower than that of the others,
the display quality of Fig. 5(d) is superior to that of Fig. 5(b)
and (c) because all of the black secret pixels can be recovered.
Because a cipher-text can be made in smaller (or thin) fonts
the secret image can have higher capacity in a given area or a
dealer can write the cipher-text in a smaller area. The above
example proves the effectiveness of the blackness constraint in
the proposed optimization model. These results also indicate
that the blackness can be more important than contrast in the
visual quality of a VC scheme.
VI. CONCLUSION
In this study, we propose a weak visual cryptogra-
phy scheme for GASs using the optimization technique.
LEE AND CHIU: IMAGE SIZE INVARIANT VC FOR GASs 3841
(a) (b)
(c) (d)
Fig. 5. A part of implementation results for access structure 10, (a) the secret
image (320 320 pixels, 192DPI), (b) the recovered image for set {2, 3, 5}
(Model A, contrast = 1/10, blackness = 4/5), (c) the recovered image
for set {2, 3, 5} (Model B, = 1/10, = 9/10), (d) the recovered image for
set {1, 3, 4, 5} (Model C, = 1/12, = 1).
The proposed model for SIVCSs eliminates the disadvantages
of the pixel-expansion problem from which conventional VC
scenarios suffer. Our method guarantees the blackness of black
secret pixels for VCSs and improves the display quality of
the worst-case image. The experimental results show that our
approach performs better than those previously proposed in
terms of the display quality of the recovered image, which
includes the controllable blackness for black secret pixels and
maintenance of the same aspect ratio as that of the original
secret image.
The major contributions of this work include the following
three: First, this is the rst solution for weak SIVCS for
GASs subject to controllable blackness of black secret pixels.
Second, we formulate the construction problem of the SIVCS
for GASs as a mathematical optimization problem such that
the problem can be solved by using optimization techniques.
Third, the proposed method is a general and systematic
approach that can be applied to any VC schemes without
individually redesigning codebooks or basis matrices.
ACKNOWLEDGMENT
Hereby, the authors appreciate the anonymous reviewers for
their valuable comments.
REFERENCES
[1] M. Naor and A. Shamir, Visual cryptography, in Advances in Cryp-
tology, vol. 950. New York, NY, USA: Springer-Verlag, 1995, pp. 112.
[2] J. Weir and W. Yan, A comprehensive study of visual cryptography,
in Transactions on Data Hiding and Multimedia Security V (LNCS),
vol. 6010. New York, NY, USA: Springer-Verlag, 2010, pp. 70105.
[3] C. N. Yang, New visual secret sharing schemes using probabilistic
method, Pattern Recognit. Lett., vol. 25, no. 4, pp. 481494, 2004.
[4] R. Ito, H. Kuwakado, and H. Tanaka, Image size invariant
visual cryptography, IEICE Trans. Fundam., vol. E82-A, no. 10,
pp. 21722177, 1999.
[5] P. L. Chiu and K. H. Lee, A simulated annealing algorithm for general
threshold visual cryptography schemes, IEEE Trans. Inf. Forensics
Security, vol. 6, no. 3, pp. 9921001, Sep. 2011.
[6] G. Ateniese, C. Blundo, A. D. Santis, and D. R. Stinson, Visual
cryptography for general access structures, Inf. Comput., vol. 129, no. 2,
pp. 86106, 1996.
[7] C. S. Hsu and Y. C. Hou, Goal-programming-assisted visual cryp-
tography method with unexpanded shadow images for general access
structures, Opt. Eng., vol. 45, no. 9, pp. 097001-1097001-10, 2006.
[8] C. S. Hsu, S. F. Tu, and Y. C. Hou, An optimization model for
visual cryptography schemes with unexpanded shares, in Foundations
of Intelligent Systems (LNAI), vol. 4203. New York, NY, USA: Springer-
Verlag, 2006, pp. 5867.
[9] F. Liu, C. Wu, and X. Lin, Step construction of visual cryptography
schemes, IEEE Trans. Inf. Forensics Security, vol. 5, no. 1, pp. 2738,
Mar. 2010.
[10] A. Adhikari, T. K. Dutta, and B. Roy, A new black and white visual
cryptographic scheme for general access structures, in Progress in
Cryptology (LNCS), vol. 3348. New York, NY, USA: Springer-Verlag,
2004, pp. 399413.
[11] L. A. MacPherson, Grey level visual cryptography for general access
structures, M.S. thesis, School of Comput. Sci., Univ. Waterloo, Water-
loo, ON, Canada, 2002.
[12] K. H. Lee and P. L. Chiu, An extended visual cryptography algorithm
for general access structures, IEEE Trans. Inf. Forensics Security, vol. 7,
no. 1, pp. 219229, Feb. 2012.
[13] F. Liu, C. K. Wu, and X. J. Lin, A new denition of the contrast
of visual cryptography scheme, Inf. Process. Lett., vol. 110, no. 7,
pp. 241246, 2010.
[14] C. Blundo and A. De Santis, Visual cryptography schemes with
perfect reconstruction of black pixels, Comput. Graph., vol. 22, no. 4,
pp. 449455, 1998.
[15] T. Hofmeister, M. Krause, and H. U. Simon, Contrast-optimal k out of
n secret sharing schemes in visual cryptography, Theoretical Comput.
Sci., vol. 240, no. 2, pp. 471485, 2000.
[16] H. Koga, A general formula of the (t, n)-threshold visual secret sharing
scheme, in Advances in Cryptology, Asiacrypt. New York, NY, USA:
Springer-Verlag, 2002, pp. 328345.
[17] Y. W. Chow, W. Susilo, and D. S. Wong, Enhancing the perceived visual
quality of a size invariant visual cryptography scheme, in Information
and Communications Security (LNCS), vol. 7618. New York, NY, USA:
Springer-Verlag, 2012, pp. 1021.
[18] H. B. Zhang, X. F. Wang, W. H. Cao, and Y. P. Huang, Visual cryp-
tography for general access structure using pixel-block aware encoding,
J. Comput., vol. 3, no. 12, pp. 6875, 2008.
Kai-Hui Lee received the Ph.D. degree in electronic
engineering from the National Taiwan University of
Science and Technology, Taipei, Taiwan, in 2002.
He is a Professor with the Department of Computer
Science and Information Engineering, Ming Chuan
University, Taipei.
His current research interests include visual cryp-
tography, wireless networks, and network resource
managements.
Pei-Ling Chiu received the Ph.D. degree in informa-
tion management from the National Taiwan Univer-
sity, Taipei, Taiwan, in 2007. She is a Professor with
the Department of Risk Management and Insurance,
Ming Chuan University, Taipei.
Her current research interests include visual cryp-
tography, wireless sensor networks, and optimizing
technologies.