Multimedia Tools and Applications

https://doi.org/10.1007/s11042-018-6544-x

High capacity reversible data hiding using 2D

parabolic interpolation

Ahmad Shaik 1 & Thanikaiselvan V 1

Received: 23 October 2017 / Revised: 8 May 2018 / Accepted: 15 August 2018

# Springer Science+Business Media, LLC, part of Springer Nature 2018

Abstract
Reversible data hiding (RDH) plays a key role in ensuring security of communications. Recent
studies have revealed that RDH techniques based on interpolation are gaining popularity
because of the ease of implementation and support for high embedding capacity. From recent
literature, it is noted that the one-dimensional (1D) parabolic interpolation-based data hiding
technique is suitable for high embedding capacity applications. This study aims to implement a
high capacity RDH technique using a new two-dimensional (2D) parabolic interpolation and a
novel embedding technique that is suitable for interpolation-based techniques. The results of
this study showed that the proposed 2D parabolic interpolation maximized the utilization of the
cover image pixels in up-sampling. The technique produced interpolated images of a higher
quality compared to the 1D parabolic interpolation technique. Moreover, the proposed data
hiding technique exploited the local redundancy and obtained the high embedding capacity
with an appreciable image quality better than other state-of-the-art techniques. The results of
this study support the view that the proposed parabolic interpolation has the potential to
address security issues in high embedding capacity applications.

Keywords Parabolic interpolation . Reversible data hiding . Image interpolation .

Interpolation-based RDH

1 Introduction

Reversible Data Hiding (RDH) is an important invention in the field of data security. The idea
behind the RDH system is to exchange the secret data by hiding it in a meaningful cover file

* Thanikaiselvan V
[email protected]

Ahmad Shaik
[email protected]

1
Department of Communication Engineering, School of Electronics Engineering (SENSE), VIT
University, Vellore, Tamilnadu -632014, India
 Multimedia Tools and Applications

while transferring over the public network to bypass the unwanted eavesdropping. The RDH
algorithms have been mainly used in medical and defense applications [11]. Especially image
based RDH techniques are suitable for medical applications to hide large amounts of patient
personal and diagnostic information inside the medical images. It helps to minimize the
documents required and also it promotes easy sharing of medical data among the doctors for
a better diagnosing of the patient condition.
In RDH techniques, according to the secret data, the cover image pixels are slightly
modified. This process is called as data embedding, and the modified cover image is called
as the marked image or stego image. The visual quality of the stego image highly correlates
with the original cover image. Hence the stego image successfully avoids the unwanted
attention. At the receiver end, the secret data and the cover image are retrieved from the stego
image using the extraction process.
Recently, a few researchers have addressed the issue using interpolation techniques. In
digital image processing, the interpolation techniques play a vital role. Hence using latter
techniques in RDH is a justifiable step to advance the embedding capacity [4]. In this
approach, at first, the cover image is interpolated using one of the standard interpolation
techniques. Afterward, interpolated pixels are altered for data hiding, and the original pixels
are kept unchanged furthermore used as reference pixels for interpolation and data hiding.
During the extraction, the secret data is retrieved from the stego image pixels and the original
cover image is restored by down-sampling the stego image. The detailed block diagram of the
interpolation-based RDH technique is shown in Fig. 1.
Various approaches have been put forward to implement interpolation-based RDH. In that, Jung
and Yoo [5] proposed a pioneering step in a method based on the Neighbor Mean Interpolation
(NMI). The results of their study infer that the interpolation-based RDH techniques can hide larger
payloads compared to other spatial techniques like difference expansion (DE) [13], histogram
shifting (HS) [9] and combination of DE and HS [12]. As an extension to it, Lee and Huang
proposed a technique using the interpolation by neighboring pixels [6]. Similarly, Hu and Li
proposed the interpolation by maximizing the difference values between neighboring pixels [3].

Interpolated Image Stego Image

Secret Data
Cover Image

Interpolation Data
Embedding

Common Channel

Down Sampling

Cover Image Data

Extraction

Secret Data Stego Image

Fig. 1 Block diagram of the proposed data hiding system

Multimedia Tools and Applications

Luo et al. [8] implemented an additive interpolation error expansion scheme by expanding
the errors between the original and the interpolated pixels. Hong and Chen [2] proposed image
interpolation and reference pixel distribution mechanism (RPDM) based data hiding. This
technique divides the cover image into the smooth and complex blocks and conceals the data
in smooth region only. As an extension to Hong and Chen [2] technique, Lu et al. [7] proposed
the DE, HS and interpolation-based approach to hide secret data. Later, Govind and Wilscy [1]
developed directional interpolation-based DE technique to improve the Lu et al. [7] technique.
Furthermore, Zang et al. [14] proposed a novel high capacity parabolic interpolation-based
RDH. In this technique, the cover image is interpolated using a one-dimensional (1D)
parabolic interpolation. This scheme offered high embedding capacity with decent image
quality compared to [5, 6], but the work has been only limited to 1D interpolation. As a
result, it failed to utilize the spatial redundancy of the cover image properly.
In their seminal article Zang et al. [14] suggested that parabolic interpolation-based data
hiding is a reliable approach. And it has a potential to produce high quality interpolated
images, useful for medical applications. Hence this study aimed to implement a high capacity
RDH technique using a new two-dimensional (2D) parabolic interpolation and a novel
embedding technique that is suitable for interpolation-based techniques. The utilization of a
more number of original pixels in the proposed data hiding process improves the data hiding
capacity while maintaining the good visual quality.
The rest of the paper is organized as follows; Section 2 gives a brief overview about the
Zang et al. [14] 1D interpolation-based data hiding. The proposed 2D parabolic interpolation-
based data hiding is described in Section 3. Section 4 deals with the results and discussion,
followed by conclusions in Section 0.

2 1D interpolation-based data hiding

Zhang et al. [14] proposed a 1D parabola-based RDH technique. In this study, the cover image
I of size M × N is up-sampled with a scaling factor 2 to obtain the interpolated image IP of size
2 M × 2 N. Then, the interpolated pixels are used to hide the secret data by utilizing the
original cover image pixels as reference. The extraction process retrieves the secret data from
the interpolated stego pixels and then the cover image is restored by down-sampling.
The interpolation process begins by filling the cover image pixels into IP using Eq. 1.

I P ð2i þ 1; 2j þ 1Þ ¼ I ði; jÞ ð1Þ

where i (0 ≤ i ≤ M-1), and j (0 ≤ j ≤ N-1) represent the row and column indexes of I
respectively. This operation stores one-fourth of IP with original cover image pixels while
the remaining pixels are available for estimation. Here, the IP is classified into four parts
according to index values i and j viz., even-even, even-odd, odd-even and odd-odd. Out of the
four parts, the pixels with odd-odd index values are already filled with the cover image pixels
and remaining pixels need to be calculated.
Let us consider a block of five consecutive pixels from IP either in a horizontal, vertical or
diagonal direction so that three alternate pixels are known and the other two are unknown.
Figure 2 represents a sample 1D interpolation block, where the shaded pixels are known pixels
and the remaining are unknown. Substitution of known pixel indexes (−2, 0 and 2) and their
respective intensities (g−2, g0 and g2) in 1D parabola (Eq. 2) gives a set of three equations in
 Multimedia Tools and Applications

Fig. 2 1D parabola interpolation -2 -1 0 1 2

block

g-2 g-1 g0 g1 g2

terms of parabola parameters a2, a1 and a0.

gx ¼ a2 x2 þ a1 x þ a0 ð2Þ

The index x (−2 ≤ x ≤ 2) and the pixel intensity gx (0 ≤ gx ≤ 255) values are represented in
(x, gx) format. A solution to these sets of equations determines the parabola which passes
through the points (−2, g−2), (0, g0) and (2, g2). From this, the unknown pixel intensities g−1
and g1 are calculated by substituting the x values as −1 and 1 in Eq. 2.
According to the position of the interpolating pixel IP (i, j), the 1D interpolation block can
be used in horizontal, vertical, or diagonal directions. If a selected pixel satisfies the 1D
interpolation block criteria in multiple ways, we can generate the multiple parabolas followed
by multiple estimations. In that case, the average of all estimations is taken as the final value
(h1). For the interpolating image IP, the number of 1D parabolas (U) that can be generated for
each pixel is shown in Table 1.
To improve the local redundancy of the interpolated image, the final interpolating pixel IP (i,
j) value is determined as a combination of the average parabolic interpolation value h1 and the
weighted average of the surrounding known pixel value h2. The mathematical representation of
the final value calculation is shown in Eq. 3, where k (0 ≤ k ≤ 1) is the controlling parameter.

I P ði; jÞ ¼ kh1 þ ð1−k Þh2 ð3Þ

In weighted average calculation, the Euclidean distance between the interpolating pixel IP (i, j)
and the known neighbor pixels is considered as the weightage (1 ≤ m ≤ V) for pixel intensities,

Table 1 1D and 2D interpolation comparison with respect to U and V

For image IP of size 2 M × 2 N 1D interpolation 2D interpolation

U V U V

If (i, j) = (1, 2), (2, 1), (1, N-2), (2, 2 N-1), (2 M-2, 1), 1 2 1 2
(2 M-1, 2), (2 M-1, 2 N-2) and (2 M-2, 2 N-1)
(2, 2), (2, 2 N-2), (2 M-2, 2) and (2 M-2, 2 N-2) 1 4 1 4
If i = 1 or 2 M-1 and j is even (4 ≤ j ≤ 2 N-4) 2 2 2 2
If j = 1 or 2 N-1 and i is even (4 ≤ i ≤ 2 N-4)
If i = 2 or 2 M-1 and j = 3 or 2 N-3 1 6 2 2
If i = 3 or 2 M-3 and j = 2 or 2 N-1
If i = 2 or 2 M-2 and j is odd (5 ≤ j ≤ 2 N-5) 1 6 3 2
If j = 2 or 2 N-2 and i is odd (5 ≤ i ≤ 2 N-5)
If i = 2 or 2 M-2 and j is even (4 ≤ j ≤ 2 M-4) 2 4 2 4
If j = 2 or 2 N-2 and i is even (4 ≤ i ≤ 2 N-4)
If i is odd (3 ≤ i ≤ 2 N-3) and j is even (4 ≤ j ≤ 2 N-4) 2 6 4 2
If i is even (4 ≤ i ≤ 2 N-4) and j is odd (3 ≤ j ≤ 2 N-3)
If i = 4 or 2 M-4 and j is odd (5 ≤ j ≤ 2 N-5) 2 6 6 2
If j = 4 or 2 N-4 and i is odd (5 ≤ j ≤ 2 M-5)
If i is even (4 ≤ i ≤ 2 M-4) and j is even (4 ≤ j ≤ 2 M-4) 4 4 4 4
Multimedia Tools and Applications

which is represented in Eq. 4. Here, #m (1 ≤ m ≤ V) and wm represent the mth original pixels
and their respective weight.
V
#m
h2 ¼ ∑ =wm ð4Þ
m¼1
V

∑ 1
wm
m¼1

Here, V represents the total number of known pixels used in weighted average. If the i value is odd
and the j value is even, calculate the weighted average using the six known pixels (#1 to #6), as
shown in the Fig. 3(a). Similarly, for the remaining two possibilities viz., even-even and even-odd,
the known pixels are shown in Fig. 3(b) and (c) respectively. If the pixel is at any of the borders of
the image, only the immediate neighbors are considered. For interpolating image IP, the V values
according to the pixel position are shown in Table 1. After performing all pixel calculations, fill
the first row and first column of the IP with their adjacent row and column values respectively.
After the 1D parabolic interpolation, the secret data is embedded in the interpolated pixels
by using the original cover image pixels as reference. In the data embedding process, a block
of 5 × 5 pixels is taken from IP at a time so that nine out of 25 pixels are original cover image
pixels and the remaining are interpolated pixels. Then, each block is further classified into
eight sequences each of size 1 × 5 so that the 1st, 3rd, and 5th pixels are the original pixels and
the remaining two are interpolated pixels. Thus, 16 pixels in all are used in each block for
hiding data and the remaining nine pixels are used as reference.
Now, consider one interpolated pixel p from a sequence. Calculate the difference d between
the interpolated value p and the average value of the neighboring pixels. Here a larger|d| implies
a big change in the pixel values and small|d|implies a small change. From the experiments, it is
noted that the images with fewer textures have smaller differences than the images with more
texture. Large|d|values make a significant impact on the stego image quality and embedding
capacity. To control the influence of|d|, a threshold T is employed. Now, the interpolated pixels
that satisfy the condition|d| ≤ T are used for embedding. If the d value satisfies the predefined
threshold T condition (|d| ≤ T), then calculate the number of secret bits v that can be embedded
into the pixel using Eq. 5. If the threshold condition is not met, ignore the pixel.
v ¼ blog2 ðjdj þ 1Þc ð5Þ

Embed the secret data by adding or subtracting the decimal equivalent of v number of secret
bits to the interpolated pixel. The addition or subtraction operation is selected according to the

#1 #2

#1 #2 #1 #2 #3

#3 IP(i, j) #4 IP(i, j) IP(i, j)

#3 #4 #4 #5 #6
#5 #6

(a) odd-even (b) even-even (c) even-odd

Fig. 3 Weighted average neighboring pixels
 Multimedia Tools and Applications

coefficient a2 of x2 using Eqs. 6 and 7. Here, two predefined parameters (non-negative

integers) s and l are used to control the embedding capacity.
In a parabola, the coefficient of x2 dictates the opening direction of the parabola. If a2 > 0
the parabola has an upward opening; if a2 < 0 means downward opening; and if a22 = 0 means
parabola forms a straight line. In data embedding, if a2 > 0, s + v number of secret bits
equivalent to decimal number n1 are added to the pixel p; if a2 < 0, s + v number of secret bits
equivalent to decimal number n1 are subtracted from the pixel p.

p þ n1 if a2 > 0
q¼ ð6Þ
p−n1 if a2 < 0

If a2 = 0, l secret bits equivalent to decimal number n2 are embedded using Eq. 7, where B1 is
the average of the nine original pixels in 5 × 5 interpolation block.

p þ n2 if jp þ n2 −B1 j ≤ jp−n2 −B1 j
q¼ ð7Þ
p−n2 if otherwise

This addition and subtraction based on a2 ensures that the stego pixel intensity q remains close to the
mean and does not exceed the boundaries. As a result, if the stego pixel is smaller than 0, the secret
bits will be added to the interpolated pixel to get the stego image pixel within the prescribed range
and if the pixel is bigger than 255, the secret bits will be subtracted from the interpolated pixel to get
the stego pixel within range. Likewise, for each sequence, two interpolated pixels are used to hide
data. This procedure is repeated until all the pixels have been processed and the stego image IS is
obtained.
Since the original cover image pixels are unchanged, the interpolated image IP can be calculated
using a parabolic interpolation. For the selected interpolated image pixel p, calculate the difference d.
If the d value satisfies the threshold condition (|d| ≤ T), find the value of v using Eq. 5. Calculate the
secret bits equivalent to decimal number d′ using Eq. 8. If a2 ≠ 0, restore the secret bits by converting
the d′ value into s + v number of binary bits and if a2 = 0, restore the secret bits by converting the d′
value into l number of binary bits.
0
d ¼ jp−qj ð8Þ

Perform the same operation on the remaining stego image pixels and retrieve the secret data. Finally,
restore the original cover image by down-sampling the stego image.

3 Proposed data hiding

The proposed methodology bears a close resemblance to the one proposed by Zhang et al.
[14]. But it differs from the existing technique in two aspects, viz., the block size of the 1D
interpolation and the number of original neighboring pixels considered at data embedding.

3.1 Proposed 2D parabolic interpolation

In the 2D interpolation, initially, the interpolating image IP is filled with original cover image
pixels according to Eq. 1.

gðx; yÞ ¼ a22 x2 y2 þ a21 x2 y þ a20 x2 þ a12 xy2 þ a11 xy þ a10 x þ a02 y2 þ a01 y þ a00 ð9Þ
Multimedia Tools and Applications

Let consider an image block of size 5 × 5 from IP such that nine out of the 25 pixels are known
and remaining are unknown. A generalized pixel block is shown in Fig. 4, where x, y and gx,y
represent horizontal index (−2 ≤ x ≤ 2), vertical index (−2 ≤ y ≤ 2) and pixel intensity (0
≤ gx,y ≤ 255) respectively. Here, the shaded positions viz., (−2, −2), (−2,0), (−2, 2), (0, −2), (0,
0), (0, 2), (2, −2), (2, 0) and (2, 2) represent the known pixel intensities viz., g-2,-2, g-2,0, g-2,2,
g0,-2, g0,0, g0,2, g2,-2, g2,0 and g2,2 respectively. Now, the substitution of these nine known pixel
values in 2D parabola equation (Eq. 9) gives a set of nine linear equations in terms of parabolic
parameters a22, a21, a20, a12, a11, a10, a02, a01 and a00. The matrix form of these sets of
equations is represented in Eq. 10.
2 3 2 3 2 32 3
g −2;−2 gð−2; −2Þ 16 −8 4 −8 4 −2 4 −2 1 a22
6 g −2;0 7 6 gð−2; 0Þ 7 6 0 0 4 0 0 −2 0 0 1 7 6 7
6 7 6 7 6 76 a21 7
6 g −2;2 7 6 gð−2; 2Þ 7 6 16 8 4 −8 −4 −2 4 2 1 7 6 a20 7
6 7 6 7 6 76 7
6 g 0;−2 7 6 gð0; −2Þ 7 6 0 0 0 0 0 0 4 −2 1 76 7
6 7 6 7 6 76 a12 7
6 g 0;0 7 ¼ 6 g ð0; 0Þ 7 ¼ 6 0 0 0 0 0 0 0 0 17 6 a11 7 ð10Þ
6 7 6 7 6 76 7
6 g 0;2 7 6 g ð0; 2Þ 7 6 0 0 0 0 0 0 4 2 17 6 7
6 7 6 7 6 76 a10 7
6 g 2;−2 7 6 gð2; −2Þ 7 6 16 −8 7
4 8 −4 2 4 −2 1 76 a02 7
6
6 7 6 7 6 7
4 g 2;0 5 4 g ð2; 0Þ 5 4 0 0 4 0 0 2 0 0 1 54 a01 5
g 2;2 g ð2; 2Þ 16 8 4 8 4 2 4 2 1 a00

Accordingly, the unknown parameters viz., a22, a21, a20, a12, a11, a10, a02, a01 and a00 can be
measured by using Eq. 11.

Calculation of these parameters gives a 2D parabola which passes through the known pixel
values. From this, the unknown pixel intensities viz., g-2,-1, g-2,1, g-1,-2, g-1,-1, g-1,0, g-1,1, g-1,2, g0,-
1 g0,1, g1,-2, g1,-1, g1,0, g1,1, g1,2, g2,-1 and g2,1 are predicted by substituting their respective x and
y values, viz., (−2, −1), (−2, 1), (−1, −2), (−1,-1), (−1, 0), (−1, 1), (−1, 2), (0, −1), (0, 1), (1,
−2), (1, −1), (1, 0), (1, 1), (1, 2), (2, −1) and (2, 1) into the Eq. 9.
If a selected pixel IP (i, j) satisfies the 2D interpolation block criteria in multiple ways, we
can generate the multiple 2D parabolas followed by multiple estimations. In that case, the
 Multimedia Tools and Applications

Fig. 4 2D parabola interpolation -2 -1 0 1 2

block

-2 g-2,-2 g-2,-1 g-2,0 g-2,1 g-2,2

-1 g-1,-2 g-1,-1 g-1,0 g-1,1 g-1,2

0 g0,-2 g0,-1 g0,0 g0,1 g0,2

1 g1,-2 g1,-1 g1,0 g1,1 g1,2

2 g2,-2 g2,-1 g2,0 g2,1 g2,2

average of all estimations is taken as the final value. For interpolating image IP, the number of
2D parabolas (U) that can be generated for each pixel is shown in Table 1.

3.1.1 Interpolated pixel calculation

The final interpolated pixel IP (i, j) is a combination of the average parabolic interpolation
value h3 and the normal average of the known pixel value h4, as shown in Eq. 12. Equation 12
is obtained by replacing h1 and h2 in Eq. 3 with h3 and h4 respectively.
I P ði; jÞ ¼ kh3 þ ð1−k Þh4 ð12Þ

In normal average h4 calculation (Eq. 13), the immediate neighbors of IP (i, j), which belong to
the original cover image are taken into consideration. Figure 5 shows the neighboring pixels
according to the row and column index values used in average calculation. The number of
known pixels (V) available for each interpolating pixel is represented in Table 1. In Eq. 13, #m
(1≤m≤V) represents the original cover image pixels.
 
1 V
h4 ¼ ∑ #m ð13Þ
V m¼1

#1 #2 #1

#1 IP(i, j) #2 IP(i, j) IP(i, j)

#3 #4 #2

(a) odd -even (b) even-even (c) even-odd

Fig. 5 Normal average neighboring pixels
Multimedia Tools and Applications

An example of the proposed 2D parabolic interpolation with k as 0.7 is shown in Fig. 6. Here,
Fig. 6(a) denotes the cover image of size 5 × 5; Fig. 6(b) denotes the interpolated image of size
10 × 10. Impact of different k values are presented in Section 4. In Fig. 6(b) the pixels with
shade are the original cover image pixels, and others are interpolated pixels.

3.2 Data embedding

In data embedding, the interpolated image IP pixels are scanned in raster scan order consid-
ering each pixel at a time. The detailed description of data hiding is as follows:
Let us consider the interpolated cover image pixel p, secret data b, and the stego image
pixel q. Calculate the h4 value for the pixel p using Eq. 13 and then calculate the difference d
using Eq. 14.
d ¼ p−bh4 c ð14Þ
where ⌊ ⌋ is the floor function. If the d value satisfies the predefined threshold T condition
(|d| ≤ T), the number of bits v that can be embedded in that pixel p can be calculated using Eq.
5.
The secret data is embedded by adding or subtracting the decimal equivalent of secret bits v
to the interpolated pixel. The addition or subtraction operation is selected according to the
coefficient a22 of x2y2 using Eqs. 15 and 16, which are modified versions of Eqs. 6 and 7,
respectively. Here, the pixels which have multiple a22 values may consider the average value
as the final value.
Here, the same parameters s and l are used to control the embedding capacity. Like the
1D interpolation, if a22 > 0, s + v number of secret bits equivalent to decimal number n3
are added to the pixel p; if a22 < 0, s + v number of secret bits equivalent to decimal
number n3 are subtracted from the pixel p. This addition and subtraction according to a22
ensures that the stego pixel q remains close to the mean and does not exceed the

0 1 2 3 4 5 6 7 8 9

0 165 165 166 166 167 167 167 167 167 167

1 165 165 166 166 167 167 167 167 167 167
0 1 2 3 4 2 165 165 165 166 167 168 168 168 167 166
0 165 166 167 167 167 3 165 165 165 165 167 168 168 168 167 166
1 165 165 168 168 166 4 166 166 164 163 165 166 166 141 166 166
2 167 162 164 164 166 5 167 167 164 162 163 164 164 164 165 166
3 159 162 160 162 161 6 163 163 162 162 162 162 162 138 163 163
4 157 161 162 159 162 7 159 159 161 162 161 160 161 162 162 161

8 157 157 160 161 161 160 160 161 161 161

9 157 157 159 161 162 162 160 159 160 162

(a) Input image of size 5×5 (b) Interpolated image of size 10×10
Fig. 6 2D parabolic interpolation example
 Multimedia Tools and Applications

boundaries.

p þ n3 if a22 > 0
q¼ ð15Þ
p−n3 if a22 < 0

If a22 = 0, l secret bits equivalent to decimal number n4 are embedded using Eq. 16, where B2 is
the average of the nine original pixels in the 5 × 5 interpolation block. If multiple parabolas are
used in the interpolation, consider the average of all means as the B2 value.

p þ n4 if jp þ n3 −B2 j ≤ jp−n3 −B2 j
q¼ ð16Þ
p−n4 if otherwise

This procedure is repeated until all the pixels have been processed and the stego image IS is
obtained.

3.2.1 Example

Consider Fig. 7 which is a collection of three interpolated pixels IP(1,2), IP(2,1) and IP(4,2)
taken from Fig. 6(b) to illustrate the data embedding process. As shown in the Fig. 7(a), for
IP(1,2), the original neighboring pixels are 165 and 166. Now we take an average of the
neighboring pixels of IP(1,2) and denote it as h4 = 165. Because the pixel can support 2D
interpolation block criteria in only one way, we can generate only one 2D parabola. Now, we
discuss the calculations for a22, B2, and h4 using the same parameters as mentioned in the example
given in Section 3.1.1. Let us assign the values T = 3, s = 1 and l = 1. The value of a22 can be
calculated from 2D parabola to be 0.0156. The difference between IP(0,1) and h4 can be calculated
as|d| = 1 and v = 1. Since a22 > 0 the parabola has an upward opening. So, 1 + 1 = 2 bits can embed

1 1 2 3

1 2 3 1 165 3 165 165 165

1 165 166 166 2 165 4 166 164 163

3 165 5 167 164 162

(a) IP(1,2) (b) IP(2,1) (c) IP(4,2)

1 1 2 3

1 2 3 1 165 3 165 165 165

1 165 169 166 2 165 4 166 163 163

3 165 5 167 164 162

(d) IS(1,2) (e) IS(2,1) (f) IS(4,2)

Fig. 7 Data embedding example
Multimedia Tools and Applications

into IP(1,2). Let the two bits be B11^ and its decimal equivalent be B3^. This value is then added to
the interpolated pixel to get the stego image pixel IS(1,2) = 169, which is shown in Fig. 7(d).
From Fig. 7(b), we can infer that for a pixel IP(2,1), h4 = 165,|d| = 0 and v = 0. Here, we can
generate only one 2D parabola. Hence, we get only one value of a22 = 0.0156. Since a22 > 0
the parabola has an upward opening. So, 1 + 0 = 1 bits can embed into IP(2,1). Let the one
secret bit be B0^ and its decimal equivalent also be B0^. This value is then added to the
interpolated pixel to get the stego image pixel IS(2,1) = 165, which is shown in Fig. 7(e).
Similarly, for a pixel IP(4,2), the average of four original pixels h2 = 164,|d| = 0 and v = 0. Here,
the pixel supports the 2D interpolation block criteria in two ways so that we get two different
parabolas. The values of a22 as calculated from 2D parabolas are 0.0156 and − 0.2500. The
average value can be treated as the final value, which is −0.1172. Since a22 < 0, the parabola
has a downward opening. So, 1 + 0 = 1 bits can embed into IP(4,2). Let the two bits be B1^ and
its decimal equivalent be B1^. This value is then subtracted from the interpolated pixel to get
the stego image pixel IS(4,2) = 163, which is shown in Fig. 7(f).

3.3 Data extraction

In the data extraction, the stego image pixels are scanned one at a time using the raster scan
approach. The secret data is extracted from the stego image pixels by using the original cover
image pixels. Because the original cover image pixels are unchanged, the interpolated image IP
can be calculated by using 2D parabolic interpolation.
For a stego pixel q, let us consider the interpolated pixel p from IP and then calculate the
average value h4 using Eq. 13. After that, calculate the difference d between p and h4 using Eq.
14. If the|d|value satisfies the condition|d| ≤ T, find the value of v using Eq. 5 and calculate the
difference d′ between the stego pixel q and the interpolated pixel p using Eq. 8. If a22 ≠ 0,
restore the secret bits by converting the d′ value into s + v number of binary bits; and if a22 = 0,
restore the secret bits by converting the d′ value into l number of binary bits.
Repeat the same operation on the remaining stego image pixels. After extraction, concat-
enate the restored bits to get the complete secret data. Finally, retrieve the original cover image
by down-sampling the stego image.
From Table 1, it is noted that to calculate a single interpolating pixel using only parabolic
interpolation, the minimum and maximum number of original cover image pixels required are

(a) Elaine (b) Lena (c) Barbara (d) Man (e) Goldhill

(f) Baboon (g) Bridge (h) Couple (i) Peppers (j) Plane
Fig. 8 Input images
 Multimedia Tools and Applications

three and six, respectively. Similarly, in the data embedding phase, the 1D interpolation
technique uses two original neighbor pixels for the calculation of difference value. On the
other hand, in the proposed technique, we used 9 to 25 original cover image pixels for a single
pixel calculation. Moreover, we have used a higher number of reference pixels for the
calculation of d value. Hence, we obtained a better quality of the interpolated image as well
as a higher-capacity stego image compared to the 1D interpolation technique.

4 Results and discussion

The input images considered for evaluation are shown in Fig. 8. The performance of the
proposed technique has been evaluated for the stego image visual quality and the embedding
capacity. Embedding bits or bits per pixel (bpp) is a parameter used to measure the embedding
capacity of the image. The peak signal-to-noise ratio (PSNR) is used to measure the visual
image quality of the stego image. The PSNR between the cover image I and stego image I′ of
size M × N is given in Eq. 17.
 
255  255
PSNRðdBÞ ¼ 10log10 ð17Þ
MSE
M N  0
where, MSE ¼ MN 1
∑ ∑ I ði; jÞ−I ði; jÞ 2
i¼1 j¼1
To evaluate the proposed interpolation, the input images of size 512 × 512 are initially
scaled down by a factor 2. These scaled-down images are again interpolated back to their
original size by using the proposed interpolation. The PSNR between the original and the
interpolated images for different k values is shown in Table 2. According to the nature of the
image, the change in control parameter k will affect the quality of the interpolated image.
From Table 2, it is noted that most of the interpolated images maintained a relatively stable
PSNR for various k values. Also, it is interesting to see that the smooth images like Lena,
Plane, Man, Goldhill, Peppers, and Boat generated high-quality interpolated images when
k ≥ 0.8, which means that for smooth images, direct parabola interpolation worked better as
compared to the linear averaging of the neighboring pixels. Contrary to that, high-textured
images like Baboon and Couple generated high-quality images when k = 0, which means that
for rough images the linear averaging performed well. For medium smooth/rough images like
Elaine and Bridge, a high PSNR is obtained when k value is in between 0.5 and 0.6.
Figure 9 shows the four interpolated images obtained from the 2D parabolic interpolation
when k = 0.7. From the Fig. 9, it is noted that the proposed 2D parabola interpolation can
produce visually high-quality images.
The existing Zhang et al.’s one dimensional parabolic interpolation technique uses a 1 × 5
pixel block (Fig. 2) which contains three original pixels for pixel calculation whereas the
proposed technique uses a 5 × 5 pixel block (Fig. 4) which contains 9 original cover image
pixels for pixel calculation. Hence the utilization of number of pixels in interpolation is more
in the proposed interpolation technique as compared to [14]. This provides an advantage to
produce relatively high correlated interpolation pixels as compared with the one-dimensional
parabolic interpolation. Figure 10 shows the PSNR comparison between the proposed inter-
polation technique and the technique proposed by Zhang et al. [14] for different k values. From
the Fig. 10(a) and (b), it is observed that the 2D interpolation works better on smooth images
like Lena and Man as compared to its counterpart. On the other hand, for textured images like
 Multimedia Tools and Applications

Table 2 Interpolated image PSNR for different k values

K 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Baboon 22.9834 22.9827 22.9813 22.9759 22.9665 22.9544 22.9386 22.9194 22.8967 22.8716 22.8425
Boats 28.9601 29.0148 29.0721 29.1203 29.1607 29.1923 29.2145 29.2294 29.2357 29.2337 29.2233
Bridge 25.6728 25.6907 25.7054 25.7154 25.7212 25.7219 25.7194 25.7112 25.6981 25.6809 25.6584
Couple 26.8107 26.7962 26.7718 26.7402 26.7029 26.6586 26.6061 26.5497 26.4871 26.4189 26.3458
Elaine 31.4212 31.4460 31.5147 31.5604 31.5885 31.6046 31.6097 31.6039 31.5878 31.5574 31.5224
Lena 32.8746 33.0587 33.2444 33.4148 33.5723 33.7098 33.8289 33.9265 34.0012 34.0492 34.0775
Man 31.2750 31.3490 31.4222 31.4892 31.5483 31.5995 31.6419 31.6755 31.7008 31.7158 31.7259
Peppers 29.0259 29.0780 29.1358 29.1867 29.2295 29.2643 29.2918 29.3108 29.3225 29.3261 29.3237
Plane 29.4201 29.6045 29.7804 29.9447 30.0981 30.2361 30.3585 30.4619 30.5471 30.6120 30.6549
Goldhill 30.5236 30.5534 30.5832 30.6074 30.6242 30.6334 30.6349 30.6304 30.6189 30.6001 30.5758
Barbara 25.0129 24.9889 24.9612 24.9308 24.8961 24.8578 24.8147 24.7700 24.7202 24.6672 24.6117
 Multimedia Tools and Applications

Fig. 9 Interpolated images

Baboon and Bridge, both techniques delivered similar results, which can be observed in
Fig. 10(c) and (d) respectively. Overall, the 2D interpolation performed well as compared to
1D interpolation because of maximum utilization of original pixels. This finding confirms that
the usage of more pixels for interpolation gives a good visual quality.
In the proposed technique, the embedding capacity can be managed by changing the
threshold T and control parameters s and l. The small parameter values support the smaller
secret data and make minor modifications to the image. The values s = 1, l = 1, and T = 3 are
the preferable values for smaller data embedding. Four stego images obtained from the
parameter values after data embedding are shown in Fig. 11. From Fig. 11, it is revealed that
the proposed data embedding technique maintained the same visual quality as observed in Fig.
9. Interestingly, the visual appearance of the edge and texture regions maintains a high
correlation with interpolated images. The results suggest that the proposed data hiding
technique can produce a good visual quality images.

(a) Lena (b) Man

(c) Baboon (d) Bridge

Fig. 10 PSNR comparison between 1D and 2D interpolation techniques with respect to k
Multimedia Tools and Applications

Fig. 11 Stego images

Figure 12 compares the results obtained from Zhang et al. and the proposed techniques.
Here, Fig. 12(a) and (c) represent embedding capacities of Lena and Peppers images with
respect to different k values respectively. Similarly, Fig. 12(b) and (d) represent the PSNR of
Lena, and Peppers images with respect to different k values respectively. A sudden drop in the
embedding capacity at k = 0.1 is observed in the proposed technique. As shown in Fig. 12(a)
and (b), the 2D parabolic interpolation-based technique reported significantly more embedding
capacity and visual quality. From the interpolation results it can be identified that the
interpolated pixels are highly correlated with the original cover image pixels as compared
with 1D parabolic interpolation. Based on this in the data hiding process we concentrated only
on the immediate neighboring pixels to calculate the difference. This way we have optimally
utilized the original cover image pixel values to improve the interpolated cover image quality
as well as embedding capacity.

(a) Lena (b) Lena

(c) Peppers (d) Peppers

Fig. 12 EC comparison between Zhang et al. [14] and the proposed techniques with respect to k
 Multimedia Tools and Applications

Table 3 PSNR (dB) comparison with respect to low EC (bpp)

Method Lena Baboon Elaine Barbara Man Goldhill Couple Boats

Jung and EC (bpp) 0.7631 1.6309 0.8970 1.1298 0.9321 0.9689 1.0158 0.8223
Yoo [5] PSNR (dB) 31.2821 22.3715 30.8097 24.1284 29.0378 29.3624 27.7958 28.8702
Zhang EC (bpp) 0.8787 1.7849 0.8903 1.3922 0.7588 1.1108 1.0532 0.8036
et al. [14] PSNR (dB) 33.3955 22.0210 31.8881 24.6678 29.2339 30.0443 26.4409 29.0290
Proposed EC (bpp) 1.1542 2.6223 1.3074 1.2000 1.2204 1.2108 1.2139 1.2183
PSNR (dB) 33.6109 22.0606 31.2045 24.7235 29.5836 30.0481 26.3240 29.0457

Table 3 compares the performance of the proposed technique with respect to low EC. Here,
for the technique proposed by Zhang et al. [14] as well as the proposed technique in this study,
the control parameters are set to s = 1, L = 1 and T = 3. It can be seen from the images Lena,
Man, Goldhill and Boats in Table 3 that the proposed technique reported significantly more EC
and PSNR than the other two techniques. From the images, Baboon, Elaine, and Couple, it is
revealed that our technique achieved a noteworthy improvement in EC at a same visual quality.
There was no significant improvement in PSNR between 2D interpolation and NMI for rough
images like Baboon and Couple. Strong evidence of improvement was found when the input
images are smooth as compared to rough images. On average, our technique was shown to
have a 0.38 bpp better EC and 0.36 dB better PSNR as compared with the Jung and Yoo [5].
Similarly, as compared with Zhang et al. [14], the EC improved nearly 0.3 bpp. A similar
response can be found with respect to larger embedding capacity as shown in Table 4.
Fig. 13 shows the performance of the proposed data hiding technique as compared to
another technique. These tests revealed that the 2D interpolation-based data hiding supports a
high embedding capacity and generates a better visual quality as compared to 1D parabolic
interpolation. From Fig. 13(a), it is evident that our method completely outperformed Tian’s
difference expansion [13], and the techniques proposed by Tai et al. [12], Jung and Yoo [5],
Lee and Haung [6] and Zhang et al. [14] at high payloads. What is striking about the graphs in
this Fig. 13(c) is that in the Barbara image, the 1D parabolic interpolation generated a good
visual quality at low bpp because of its image features. As expected, our experiments prove
that the 2D parabolic interpolation-based data hiding can be a reliable technique for high
embedding capacity applications.
Fig. 14 represents the JPEG images of size 384 × 512 which are used in the data hiding
process [10]. Here initially we have taken the standard images and converted in to joint
photographic experts group (JPEG) standard using Matlab 2016a. These converted JPEG

Table 4 PSNR (dB) comparison with respect to high EC (bpp)

Method Lena Baboon Elaine Barbara Man Goldhill Couple Boats

Lee and Haung EC (bpp) 1.4703 2.4388 1.5900 1.8106 1.6476 1.6916 1.7596 1.4664
[6] PSNR 29.9402 21.7101 30.0864 23.6238 27.9343 28.5480 26.8605 27.7489
(dB)
Zhang et al. [14] EC (bpp) 1.7576 1.9023 1.3702 1.7120 1.4796 1.5986 1.5353 1.7312
PSNR 31.4367 21.2205 31.0234 24.0695 28.0664 28.6747 27.4470 27.4039
(dB)
Proposed EC (bpp) 1.8368 3.0684 1.8962 1.8941 1.8613 1.9304 1.9459 1.8944
PSNR 32.2445 21.0547 30.6406 24.5451 28.8264 29.3619 26.1136 28.5130
(dB)
Multimedia Tools and Applications

(a) Lena (b) Boats

(c) Barbara (d) Goldhill

Fig. 13 Performance evaluation of interpolation-based data hiding techniques with respect to PSNR and EC

images are used in the data hiding process. The embedding capacity and the visual quality of
the JPEG images are shown in Table 5. Here PSNR1 and PSNR2 represent the interpolated
cover image quality and the stego image quality respectively. Here we have performed the
image interpolation and data hiding operations with the generalized control parameter values
which are k = 0.7, s = 1, l = 1 and T = 3. From Table 5 the results are indicating that the

(a) Image1 (b) Image2 (c) Image3 (d) Image4 (e) Image5
Fig. 14 JPEG images

Table 5 Proposed data hiding results on JPEG images

Image PSNR1 (dB) Capacity (bits) Bitrate (bpp) PSNR2 (dB)

Image 1 26.4106 349,578 1.7333 25.9911

Image 2 27.4436 356,172 1.8116 26.9518
Image 3 24.2616 325,236 1.6542 24.0054
Image 4 30.4898 355,379 1.8075 29.5719
Image 5 29.2003 316,434 1.6094 28.5945
Average 26.8447 345,152 1.7480 26.3504
 Multimedia Tools and Applications

proposed technique has achieved an average capacity of 1.7480 bpp with a PSNR of
26.3504 dB. These results indicate that the proposed data hiding model has obtained a better
PSNR with less visual quality in JPEG images as compared to uncompressed images.

5 Conclusions

The primary goal of this study was to put forth a high capacity RDH technique based on 2D
parabolic interpolation. This study investigated the performance of 2D parabolic interpolation in
RDH. The results of this study indicate that the proposed interpolation technique yielded a better
quality of up-sampled images compared to other techniques. Moreover, the utilization of a higher
number of original pixels in the proposed data embedding process improved the data hiding
capacity while maintaining the good visual quality. From the results, it is noted that the proposed
data embedding technique outperformed the other state of the art techniques with respect to PSNR
and embedding capacity. The current findings add to a growing body of literature on parabolic
interpolation-based RDH techniques. A limitation of this study is that the processing time of the
proposed interpolation technique is slightly higher than the 1D interpolation. In future, studies can
examine possibilities of reducing the processing time of the stated technique.

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

References

1. Govind PVS, Wilscy M (2015) A new reversible data hiding scheme with improved capacity based on
directional interpolation and difference expansion. Procedia Comput Sci 46:491–498. https://doi.org/10.1016
/j.procs.2015.02.073
2. Hong W, Chen TS (2011) Reversible data embedding for high quality images using interpolation and
reference pixel distribution mechanism. J Vis Commun Image Represent 22:131–140. https://doi.
org/10.1016/j.jvcir.2010.11.004
3. Hu J, Li T (2015) Reversible steganography using extended image interpolation technique. Comput Electr
Eng 46:447–455. https://doi.org/10.1016/j.compeleceng.2015.04.014
4. Jung K-H (2017) A survey of interpolation-based reversible data hiding methods. Multimed Tools Appl.
https://doi.org/10.1007/s11042-017-5066-2
5. Jung K-H, Yoo K-Y (2009) Data hiding method using image interpolation. Comput Stand Interfaces 31:465–
470. https://doi.org/10.1016/j.csi.2008.06.001
6. Lee C-F, Huang Y-L (2012) An efficient image interpolation increasing payload in reversible data hiding.
Expert Syst Appl. https://doi.org/10.1016/j.eswa.2011.12.019
7. Lu T-C, Chang C-C, Huang Y-H (2014) High capacity reversible hiding scheme based on interpolation,
difference expansion, and histogram shifting. Multimed Tools Appl 72:417–435. https://doi.org/10.1007
/s11042-013-1369-0
8. Luo L, Chen Z, Chen M et al (2010) Reversible image watermarking using interpolation technique. IEEE
Trans Inf Forensics Secur 5:187–193. https://doi.org/10.1109/TIFS.2009.2035975
9. Ni ZNZ, Shi Y-QSY-Q, Ansari N, Su WSW (2006) Reversible data hiding. IEEE Trans Circuits Syst Video
Technol 16:354–362. https://doi.org/10.1109/TCSVT.2006.869964
10. Schaefer G (2003) UCID: an uncompressed color image database. Proc SPIE 472–480. doi: https://doi.
org/10.1117/12.525375
11. Shi Y, Li X, Zhang X et al (2016) Reversible data hiding: advances in the past two decades. IEEE Access 4:
3210–3237. https://doi.org/10.1109/ACCESS.2016.2573308
12. Tai W-L, Yeh C-M, Chang C-C (2009) Reversible data hiding based on histogram modification of pixel
differences. IEEE Trans Circuits Syst Video Technol 19:906–910. https://doi.org/10.1109
/TIFS.2010.2066971
Multimedia Tools and Applications

13. Tian J (2003) Reversible data embedding using a difference expansion. IEEE Trans Circuits Syst 13:890–
896. https://doi.org/10.1109/TCSVT.2003.815962
14. Zhang X, Sun Z, Tang Z, et al (2016) High capacity data hiding based on interpolated image. Multimed
Tools Appl 9195–9218. doi: https://doi.org/10.1007/s11042-016-3521-0

Ahmad Shaik is a research associate at VIT University Vellore, India. He received his B. Tech and MTech
degrees in Electronics and Communication Engineering and Digital Electronics and Communication Systems
from the JNTU University, Ananthapur in 2012 and 2015, respectively. And he is currently pursuing PhD in VIT
University Vellore, India. He is the author of two journal papers and has written a book chapter. His current
research interests include reversible data hiding and integer wavelet transforms.

DR. V. Thanikaiselvan received his Ph.D. degree in the field of Information security form VIT university,
Vellore, Tamilnadu in the year 2014. He received M. Tech in Advanced Communication Systems from SASTRA
University, Thanjavur, in 2006 and B. E in Electronics and Communication Engineering form Bharathidasan
University, Trichy, in 2002. He joined VIT University in 2006 as a Lecturer and Currently working as an
Associate Professor in Department of Communication Engineering under the School of Electronics Engineering,
VIT University, Vellore. His teaching and research interest includes Digital communication, Wireless commu-
nication, Digital signal and Image Processing, Wireless Sensor Networks and Information Security. He has
published more than 30 research articles. Currently he is guiding five Ph.D. candidates in the areas of Information
Security and Digital Image Processing. He is an active IEEE member and reviewer for Elsevier, Wiley and
springer journals.