Multimed Tools Appl DOI 10.1007/s11042-014-1934-1
Blind reliable invisible watermarking method in wavelet domain for face image watermark Himanshu Agarwal · Balasubramanian Raman · Ibrahim Venkat
© Springer Science+Business Media New York 2014
Abstract In this paper, we have combined watermarking and biometrics for possible improvement in owner identification/verification technology. We have proposed and compared wavelet based four blind invisible watermarking methods that have used face image as watermark. Two watermarking methods are based on the discrete wavelet transform (DWT) and rest two watermarking methods are based on the redundant discrete wavelet transform (RDWT). One watermarking method in each transform incorporates weighted binary coding to achieve improved reliability of extracted watermark. Other watermarking methods replace original image coefficients with face image coefficients. We have observed that DWT based watermarking methods outperform RDWT based watermarking methods. We have compared the robustness of the proposed watermarking methods against various common image processing attacks/operations. We have observed that DWT based watermarking method coupled with weighted binary coding has the best performance without attacks; peak signal to noise ratio value of watermarked image is greater than 50 dB and normalized correlation coefficient value of extracted watermark is 1 at the watermark embedding strength of 1. Moreover, the same watermarking method has the best robustness against most of the attacks. Keywords Redundant discrete wavelet transform (RDWT) · Discrete wavelet transform (DWT) · Biometrics · Watermarking · Peak signal to noise ratio (PSNR) · Normalized correlation coefficient (NC) H. Agarwal () Department of Mathematics, Indian Institute of Technology Roorkee, Roorkee 247667, India e-mail:
[email protected] B. Raman Department of Computer Science and Engineering, Indian Institute of Technology Roorkee, Roorkee 247667, India e-mail:
[email protected] I. Venkat School of Computer Sciences, Universiti Sains Malaysia, 11800 USM, Pulau Pinang, Malaysia e-mail:
[email protected]
Multimed Tools Appl
1 Introduction Digital watermarking offers effective solution for digital rights (proof of ownership, owner identification, copyright protection, media authentication etc.) of digital media [4, 60, 64]. Digital watermarking is a method that embeds some information into a digital media content using a watermark embedder to obtain a watermarked digital media content [33]. The embedded information is called watermark and the digital media content is called cover work. Later, watermark extractor can extract watermark from the watermarked digital media content for the required solution of digital rights. Depending on the required information of original data (cover work and original watermark) in watermark extractor, watermarking methods are divided into three categories, namely blind (oblivious or public), non-blind (non-oblivious or private) and semi-blind watermarking methods [15, 45]. Watermarking methods that do not require the information of original data in their watermark extractor are called blind watermarking methods. Non-blind watermarking methods require complete information of original data in their watermark extractor while, semi-blind watermarking methods need a part of information of original data in their watermark extractor. Although, developing blind watermarking methods is the most challenging, however, these methods are more attractive than non-blind and semi-blind watermarking methods, since, in many watermarking applications, original data can not be made available at watermark extractor [19, 38, 44]. Moreover, blind watermarking methods are memory efficient as these methods do not require original data in watermark extractor. In many watermarking applications, invisibility is the foremost demand, that is there should not be any perceptual difference between the watermarked content and original content, as usually the value of distorted digital media becomes low [5, 6, 43, 46]. Moreover, reliable extraction (difference between original watermark and extracted watermark should be within acceptable range (ideally no difference)) and robustness against common image processing operations i.e filtering, cropping, noise addition etc. should not effect the extracted watermark are general requirements in all the watermarking applications [5, 6, 32, 53]. Watermark can be embedded either in spatial or transform domain. Transform domain watermarking methods have several advantages over spatial domain watermarking methods [33]. Discrete wavelet transform (DWT) is a traditional transform in watermarking [34] and redundant discrete wavelet transform (RDWT) is a modern transform in watermarking [63]. Watermarking methods can be divided mainly into five classes according to the type of watermark to be embedded in cover work as follows: random binary sequence watermark [34], random Gaussian sequence watermark [29, 69], binary logo watermark [22], gray scale image watermark [21, 35] and biometric watermark (fingerprint minutiae [24], eigen-face coefficients [24, 37], iris code [62], Mel-frequency cepstral coefficients (MFCCs) [63]). Biometrics (the automatic recognition of individuals based on their physiological and/or behavioral characteristics) offers high level of security and convenience in identification/verification applications [51]. Although, biometrics are not fool proof, however, it is being successfully used in many commercial, government and forensic applications [1, 23, 25, 51, 55]. Recent successful use of biometrics in watermark is a new motivation for the researchers to deploy more commercial applications of biometric watermarking. In this paper, our aim is to develop blind-reliable-invisible-watermarking-method in wavelet domain that use face image as watermark and digital image as cover work. The extracted face can be used for related watermarking applications (for example, owner
Multimed Tools Appl
identification, tracking, fingerprinting etc.). We have used face image instead of its features (eigen-face coefficients), as face image can be adapted with all face recognition algorithms. In this paper, we have extended watermarking methods of Kundur et al. [34] and Vatsa et al. [63] for face image watermark, as a result, we have developed four watermarking algorithms namely S1 , S2 , S3 , and S4 . We have used novel weighted binary coding in two watermarking schemes that drastically improves the reliability of extracted watermark and maintains high invisibility. It is observed that S3 -a DWT based watermarking scheme with weighted binary coding is the only watermarking scheme that extracts the watermark without error and its invisibility is the highest. The performance of watermarking schemes have been studied against various image processing operations such as cropping, Gaussian filter, Gaussian noise, and salt and pepper noise. An unexpected novel observation is that S3 has better performance than RDWT (a modern transform in watermarking) based watermarking schemes in most of the studied cases. The rest of the paper is organized as follows. In Section 2, literature survey is done on related work. Watermarking algorithms and weighted binary coding are presented in Section 3. In Section 4, experiment environment and results are discussed. Finally, conclusions are drawn in Section 5.
2 Literature survey Table 1 gives a summary of biometric watermarking scheme (in biometric watermarking scheme, host image/watermark is/are a biometric trait) and Table 2 gives a summary of wavelet based watermarking schemes from different aspects. Highlights of the literature survey are as follows. 1. For biometric related applications, mostly blind watermarking schemes are popular. 2. In biometric watermarking schemes, binary sequence/binary image, features of biometric trait (iris code, Eigen-face coefficients, fingerprint minutiae, MFC coefficients) or biometric trait image (face image) are used as watermarks. 3. In most of the biometric watermarking scheme, length of watermark is small. 4. Noore et al. [47] and Kim and Lee [31] use face image as a watermark. In [47], high value of error is observed in extracted face watermark. In [31], size of watermark is small. 5. Gunsel et al. [17] is the most popular biometric watermarking scheme. It operates in spatial domain and length of watermark is very small. 6. Kundur and Hatzinakos [34] is the most popular blind watermarking scheme that use DWT and its use has been found in [13, 18, 59, 67]. 7. In [63], a semi-blind biometric watermarking scheme is proposed using the RDWT. Its modification as a blind watermarking scheme is easy. In [63], use of the DWT for biometric watermarking is criticized and use of the RDWT for biometric watermarking is encouraged.
3 Watermarking methods and weighted binary coding In this section, we propose four wavelet based blind watermarking methods namely S1 , S2 , S3 and S4 . All the proposed methods use a digital image as cover work and a face image as watermark. S1 is an extended version of Vatsa et al. [63] and S2 is similar to S1 except it
Wavelet+LSB
Wavelet
DCT
[58]
[11]
Eigen-face coefficients
Binary logo
Binary iris feature
Palm print and iris
Face
Binary logo
Blind
Blind
Blind
Non-blind
Blind
Blind
Semi-blind
–
Blind
Blind
Binary
Face + binary text
[14]
DCT+DWT
[68]
Correlation analysis
Wavelet
[47]
Blind Blind
Fingerprint features
Blind
Blind
Blind
Blind
Blind
Blind
Blind
Small
Small
Small
Large
Small
Small
Large
Small
Small
Large
Small
Small
Small
Small
Small
Small
Small
Small
Small
Small
of original –
Length of watermark
Requirement
Iris features
[52]
Fourier transform
[48]
Spatial
Fourier transform
[27]
Encrypted iris code
[31]
Wavelet
[28]
Fingerprint minutiae
Fingerprint minutiae
Eigen-face coefficients
[34]
[18]
DCT + [17]
[34]
[67]
Binary
[37]
[12, 17]
[30]
Eigen-face coefficients, fingerprint minutiae
MFCC (voice feature)
[17]
[9]
Binary iris code
DCT
Spatial, DCT, [54]
[62]
A string
Fingerprint minutiae, eigen-face coefficients
RDWT
Spatial
[24]
[63]
Wavelet
[54]
Type of watermark
[8]
Domain
Scheme
Table 1 Comparison of biometric watermarking methods from different aspects Watermark is a
Yes
No
Yes
Yes
Yes
Yes
Yes
No
No
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
Yes
Yes
No
biometric
face image
–
–
–
original image
Heavily rely on watermark and
Small size of watermark
–
DWT use is criticized
–
–
High error in extracted watermarks
Detector
–
–
–
–
–
Degradation in watermarked
Existing schemes
–
–
Other remarks
Multimed Tools Appl
Extractor
Detector
Extractor
Extractor
Detector
Extractor
Extractor
Extractor
Extractor
Extractor
Detector
Detector
Detector
Extractor
Extractor
Extractor
Extractor
Extractor
Extractor
[12]
[34]
[29]
[3]
[20]
[21]
[61]
[69]
[35, 56]
[10]
[2]
[16]
[41]
[59]
[40]
[36]
[13]
[26]
[50]
[22]
–
Detector
[65]
Extractor/detector
Schemes
Blind
Semi-blind
Semi-blind
Blind
Blind
Semi-blind
Semi-blind
Semi-blind
Non-blind
Non-blind
Non-blind
Blind
Non-blind
Blind
–
Non-blind
Blind
Blind
Non-blind
Non-blind
Semi-blind
Requirement of original
Binary image
Binary
Gray scale
Binary image
Random binary logo
Random binary sequence
Random binary sequence
Binary image
Binary sequence
Binary, gray scale image
Random Gaussian sequence
Binary
Binary, gray scale image
Function of original image
Random binary sequence
Random Gaussian sequence
Random binary sequence
Random uniform distributed sequence
Binary logo
Random Gaussian sequence
Random binary sequence
Watermark
Table 2 Comparison of watermarking methods in wavelet domain from different aspects
–
16 × 16 –
–
Very small size of watermark
8×8
SVD+DWT
32 × 32
Audio watermarking, similar to [34]
Small size of watermark
32 × 32 –
User defined watermark may not be used Similar to [34]
–
–
–
Artifacts in watermarked image
32 × 32
< 1024
Very small
–
–
–
–
biometric watermark
User defined watermark can’t be used
–
–
Modification required to embed
–
–
–
–
Other remarks
–
–
–
–
–
–
–
< Size of original image
< 236
Length of watermark
Multimed Tools Appl
Multimed Tools Appl
uses DWT instead of RDWT. We discuss weighted binary coding for S3 and S4 . Weighted binary coding converts a face image in binary sequence and reconstructs a face image from an extracted sequence of bits. S3 is an extended version of Kundur et al. [34] watermarking method. The main improvement in S3 is that it uses face image as watermark instead of binary sequence. S4 is similar to S3 except DWT is replaced by RDWT. 3.1 Extended watermarking method of [63] using RDWT: S1 Vatsa et al. [63] embedded watermark coefficients in the locations those were excluded by phase congruency. They selected embedding locations randomly from available embedding locations using a key. They used embedding locations in extraction process. Embedding locations depend on phase congruency and consequently on the original image, therefore watermarking method proposed by Vatsa et al. [63] is semi-blind. In S1 , we select randomly embedding locations from the detailed bands of the original image those are obtained by applying RDWT/DWT on it. Therefore, in S1 , embedding locations do not depend on the original image for given image size. This slight modification in embedding locations selection makes S1 a blind watermarking method while having unnoticeable effect on the quality of watermarked images and extracted watermarks. A watermarking method consists of mainly two algorithms namely embedding algorithm (watermark embedder) and extraction algorithm (watermark extractor). The details of embedding algorithm followed by extraction algorithm are as follows. 3.1.1 Embedding algorithm of S1 Input: – – –
–
Gray scale original image Io = {Io (m1 , n1 ) : m1 = 1, 2, · · · M1 ; n1 = 1, 2, · · · N1 } of size M1 × N1 . Watermark (a face image) Wo = {Wo (m2 , n2 ) : m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 } of size M2 × N2 . Embedding locations selection key ckey1 = {ckey1 (m2 , n2 ) → (i, j, k) : i ∈ {1, 2, · · · M1 }; j ∈ {1, 2, · · · N1 }; k ∈ {H, V , D}; m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 }. H , V and D represent detailed bands of the image Io . Note that ckey1 (m2 , n2 ) should be different for different (m2 , n2 )s. Watermark embedding strength α1 .
Processing: 1. Apply 1-level RDWT on the image Io to obtain Iˆ = {Iˆ(u1 , v1 , l1 ) : u1 = 1, 2, · · · M1 ; v1 = 1, 2, · · · N1 ; l1 = A, H, V , D}. The subset Iˆ A = {Iˆ(u1 , v1 , A) : u1 = 1, 2, · · · M1 ; v1 = 1, 2, · · · N1 } is corresponding to A that gives approximation part of the image and called approximation band. Similarly, subsets Iˆ H , Iˆ V , IˆD are corresponding to H , V and D respectively that give detailed part of the image and called detailed bands. ˆ 2. Define Iˆw = I. 3. Update Iˆw by embedding watermark coefficients as follows Iˆw (ckey1 (m2 , n2 )) = α1 Wo (m2 , n2 ); m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 .
(1)
4. Apply 1-level inverse RDWT, followed by floating point truncation on the updated Iˆw to obtain the watermarked image Iw .
Multimed Tools Appl
Output: Watermarked image Iw of size M1 × N1 that has same format as of Io . 3.1.2 Extraction algorithm of S1 Input: – – –
An image I = {I (m1 , n1 ) : m1 = 1, 2, · · · M1 ; n1 = 1, 2, · · · N1 } of size M1 × N1 (that may be watermarked, unmarked or attacked). Embedding locations selection key ckey1 same as defined in Section 3.1.1. Watermark embedding strength α1 same as defined in Section 3.1.1.
Processing: 1. Apply 1-level RDWT on the image I to obtain Iˆ = {Iˆ (u1 , v1 , l1 ) : u1 = 1, 2, · · · M1 ; v1 = 1, 2, · · · N1 ; l1 = A, H, V , D}. 2. Use the following formula followed by floating point truncation to extract watermark coefficients as We (m2 , n2 ) =
Iˆ (ckey1 (m2 , n2 )) ; m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 . α1
(2)
Output: Extracted watermark We = {We (m2 , n2 ) : m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 } of size M2 × N2 that has same format as of Wo . 3.2 Extended watermarking method of [63] using DWT: S2 The use of DWT changes the structure of embedding locations selection key. The details of watermarking method S2 are as follows. 3.2.1 Embedding algorithm of S2 Input: – – –
–
Gray scale original image Io of size M1 × N1 similar to defined in the Section 3.1.1. Watermark (a face image) Wo of size M2 × N2 similar to defined in the Section 3.1.1. Embedding locations selection key ckey2 = {ckey2 (m2 , n2 ) → (i, j, k), i ∈ M1 N1 1, 2, · · · 2 ; j ∈ 1, 2, · · · 2 ; k ∈ {H, V , D}; m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 }. Note that ckey2 (m2 , n2 ) should be different for different (m2 , n2 )s. Watermark embedding strength α2 .
Processing:
1. Apply 1-level DWT on the image Io to obtain Iˆ = Iˆ(u1 , v1 , l1 ); u1 = 1, 2, · · · M21 ; v1 = 1, 2, · · · N21 ; l1 = A, H, V , D . ˆ 2. Define Iˆw = I. 3. Update Iˆ w by embedding watermark coefficients as follows Iˆw (ckey2 (m2 , n2 )) = α2 Wo (m2 , n2 ); m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 .
(3)
4. Apply 1-level inverse DWT, followed by floating point truncation on the updated Iˆw to obtain the watermarked image Iw .
Multimed Tools Appl
Output: Watermarked image Iw of size M1 × N1 that has same format as of Io . 3.2.2 Extraction algorithm of S2 Input: – – –
An image I of size M1 × N1 similar to defined in Section 3.1.2. Embedding locations selection key ckey2 same as defined in Section 3.2.1. Watermark embedding strength α2 same as defined in Section 3.2.1.
Processing:
1. Apply 1-level DWT on the image I to obtain Iˆ = Iˆ (u1 , v1 , l1 ) : u1 = 1, 2, · · · M21 ; v1 = 1, 2, · · · N21 ; l1 = A, H, V , D . 2. Use the following formula followed by floating point truncation to extract watermark coefficients as follows We (m2 , n2 ) =
Iˆ (ckey2 (m2 , n2 )) ; m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 . α2
(4)
Output: Extracted watermark We of size M2 × N2 that has same format as of Wo . 3.3 Weighted binary coding for a face image Kundur et al. [34] proposed a watermarking method that embeds/extracts a sequence of bits. In this paper, we incorporate a novel weighted binary coding in Kundur et al. [34] watermarking method for utilizing face image watermark. Weighted binary coding consists of two algorithms namely, weighted binary encoding and weighted binary decoding. Weighted binary encoding converts a face image in a sequence of bits by assigning more weight to more significant bits. Weighted binary decoding is a reverse process of weighted binary encoding. It reconstructs a face image from an extracted sequence of bits by using more extracted bits for estimating more significant bits. 3.3.1 Weighted binary encoding This section discusses a novel theory for weighted binary encoding followed by algorithm for weighted binary encoding. Let Wo = {Wo (m2 , n2 ) : m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 } be an eight bit gray scale image of size M2 × N2 , and number of available embedding locations for watermarking is Nw . Assumptions: We have made following assumptions in the theory of weighted binary encoding.
A1. A2. A3. A4.
All the pixels Wo (m2 , n2 )s of Wo have equal importance. Second least significant bits (LSBs) of Wo (m2 , n2 )s are twice important than LSBs, third LSBs are three times important than LSBs and so on. Each bit of the Wo (m2 , n2 )s should be embedded at least once during watermarking, if possible. This assumption helps in avoiding the information loss of watermark during watermarking. Odd redundancy of watermark bits is better than even redundancy.
Multimed Tools Appl
Statements: Based on the above assumptions, we have made following statements.
S 1. S 2.
According to assumption A1, number of available embedding locations for each w . Wo (m2 , n2 ) is d = M2N×N 2 According to A2, number of available embedding locations for m3 th LSB of each Wo (m2 , n2 ) is mm3 , where 8 mm3 = d. (5) m3 =1
S 3.
S 4.
Number of available embedding locations for each bit of each Wo (m2 , n2 ) is a natural number. S 2 may produce mm3 as a non-natural number. Therefore, number of available embedding locations for m3 th LSB of each Wo (m2 , n2 ) is updated with R(mm3 ), where, R(mm3 ) is a natural number in a neighborhood of the mm3 . In view of assumptions A3 and A4, R(mm3 ) must satisfy the following. – – – –
|R(mm3 ) − mm3 | is minimum. R(mm 3 ) is an odd integer. R(mm3 ) = d. R(mm3 ) ≥ 1 for all m3 = 1, 2, · · · 8, if possible. Table 3 gives possible R(mm3 )s for two different values of Nw .
In our experiments, we have used the values of Nw and R(mm3 ) that have been reported in Table 3. Algorithm A1 for weighted binary encoding: The algorithm A1 converts Wo (an eight bit gray scale image of size M2 × N2 ) in binary form W. It satisfies statements S 1-S 4. It takes two inputs: Wo and Nw and returns one output: W. The steps involved in algorithm A1 are as follows. w Step 1. Compute d = M2N×N . 2 Step 2. Compute m = dm , m3 = 1, 2, · · · , 8. 3 Step 3. Find R(mm3 ) according to the statement S 4, m3 = 1, 2, · · · , 8. Table 3 R(mm3 )s for two different values of Nw . M2 × N2 = 64 × 64 m3
Nw = 3 × 256 × 256
Nw = 3 × 512 × 512
d = 48
d = 192
m = 1.33
m = 5.33
mm3
R(mm3 )
mm3
R(mm3 )
1
1.33
1
5.33
5
2
2.66
3
10.66
11
3
3.99
5
15.99
17
4
5.32
5
21.32
21
5
6.65
7
26.65
27
6
7.98
7
31.98
31
7
9.31
9
37.31
37
8
10.64
11
42.64
43
Multimed Tools Appl
Step 4.
Convert a Wo (m2 , n2 ) in weighted binary form as follows b8 b8 · · · R(8m) times b7 b7 · · · R(7m) times . Wo (m2 , n2 ) → b8 b7 · · · b1 → = W (m2 , n2 , m3 , n3 ); . . b1 b1 · · · R(m) times (6) = where, b8 b7 · · · b1 is binary representation of Wo (m2 , n2 ); W {W (m2 , n2 , m3 , n3 ) ∈ {0, 1} : m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 ; m3 = 1, 2, · · · 8; n3 = 1, 2, · · · R(mm3 )}. Note that m3 s determine significance of corresponding bits and n3 s handle redundant bits.
3.3.2 Algorithm A2 for weighted binary decoding The algorithm A2 estimates watermark We from an extracted sequence of bits Wex = {Wex (m2 , n2 , m3 , n3 ) ∈ {0, 1} : m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 ; m3 = 1, 2, · · · 8; n3 = 1, 2, · · · R(mm3 )}. It takes one input Wex and returns one output We . In the algorithm A2 , we assume that algorithm A1 was used for encoding. The steps involved in A2 are as follows.
Step 2.
Define sets Wd1 (m2 , n2 , m3 )s, m2 = 1, 2, · · · M2 , n2 = 1, 2, · · · N2 , m3 = 1, 2, · · · 8; Wd1 (m2 , n2 , m3 ) = {Wex (m2 , n2 , m3 , n3 ) ∈ {0, 1} : n3 = 1, 2, · · · R(mm3 )}. Estimate bits Wd2 (m2 , n2 , m3 )s from corresponding sets Wd1 (m2 , n2 , m3 )s
Step 3.
0 if φ0 (Wd1 (m2 , n2 , m3 )) > φ1 (Wd1 (m2 , n2 , m3 )) ; 1 if φ0 (Wd2 (m2 , n2 , m3 )) ≤ φ1 (Wd2 (m2 , n2 , m3 )) (7) where, φ0 returns number of 0s, and φ1 returns number of 1s. Reconstruct watermark coefficients We (m2 , n2 )s
Step 1.
Wd2 (m2 , n2 , m3 ) =
We (m2 , n2 ) =
8
2m3 −1 Wd2 (m2 , n2 , m3 ).
(8)
m3 =1
Step 4.
Estimate extracted watermark We of size M2 × N2 as We = {We (m2 , n2 ) : m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 }.
3.4 Extended watermarking method of [34] in DWT: S3 Watermarking method S3 expands on Kundur et al. [34] watermarking method coupled with weighted binary encoding. The main differences between S3 and watermarking method proposed by Kundur et al. [34] are as follows: (i) In embedding algorithm of S3 , weighted binary encoding converts face image in binary form; (ii) in extraction algorithm of S3 , weighted binary decoding reconstructs face image from the extracted sequence of bits. The details of watermarking method S3 are as follows.
Multimed Tools Appl
3.4.1 Embedding algorithm of S3 Input: – – –
–
Gray scale original image Io of size M1 × N1 similar to defined in Section 3.1.1. Watermark (a face image) Wo of size M2 × N2 similar to defined in Section 3.1.1. Embedding locations selection key ckey 3 = {ckey3 (m2 , n2 , m3 , n3 ) → (i, j, k);
i ∈ 1, 2, · · · M21 ; j ∈ 1, 2, · · · N21 ; k ∈ {H, V , D}; m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 ; m3 = 1, 2, · · · 8; n3 = 1, 2, · · · R(mm3 ) (Section 3.3.1)}, where, Nw = 3 × M21 × N21 . Watermark embedding strength α3 .
Processing: 1. Apply 1-level DWT on the image Io to obtain Iˆ similar to discussed in Section 3.2.1. 2. We have embedded one bit in each position of each detailed band of the image. Therefore, number of available embedding locations for watermarking is Nw = 3 × M21 × N21 . w . Compute d = M2N×N 2 3. Convert Wo in binary form W by using weighted binary encoding. ˆ 4. Define Iˆw = I. ˆ 5. Update Iw by embedding W as follows: Iˆw (ckey3 (m2 , n2 , m3 , n3 )) = d(ckey3 (m2 , n2 , m3 , n3 ))+ ((d(ckey3(m2 , n2 , m3 , n3 )) + α3 W (m2 , n2 , m3 , n3 )) mod 2α3 ), d(ckey3 (m2 , n2 , m3 , n3 )) = Iˆ(ckey3 (m2 , n2 , m3 , n3 ))+ r(ckey3 (m2 , n2 , m3 , n3 )) + α23 , r(ckey3 (m2 , n2 , m3 , n3 )) = I (ckey3 (m2 , n2 , m3 , n3 )) mod α3 ;
(9)
where, (m2 , n2 , m3 , n3 )s are according to Section 3.4.1 and mod is the general modulo operation such that quotient is an integer. 6. Apply 1-level inverse DWT followed by floating point truncation on the updated Iˆw to obtain the watermarked image Iw . Output: Watermarked image Iw of size M1 × N1 that has same format as of Io . 3.4.2 Extraction algorithm of S3 Input: – – –
An image I of size M1 × N1 similar to defined in Section 3.1.2. Embedding locations selection key ckey3 same as defined in Section 3.4.1. Watermark embedding strength α3 same as defined in Section 3.4.1.
Processing:
1. Apply 1-level DWT on the image I to obtain Iˆ similar to discussed in Section 3.2.2. 2. Use the following formula to extract watermark bits Wex (m2 , n2 , m3 , n3 )s as follows Wex (m2 , n2 , m3 , n3 ) = roundoff2α3 (Iˆ (ckey3 (m2 , n2 , m3 , n3 )) mod 2α3 ),
(10)
Multimed Tools Appl
where,
roundoff2α3 (x) =
0 if x < α3 , 1 if x ≥ α3
(11)
and (m2 , n2 , m3 , n3 )s are according to Section 3.4.2. 3. Reconstruct watermark We from extracted bits Wex (m2 , n2 , m3 , n3 )s by using weighted binary decoding. Output: Extracted watermark We of size M2 × N2 that has same format as of Wo . 3.5 Extended watermarking method of [34] in RDWT: S4 The use of RDWT changes the structure of embedding locations selection key. The details of watermarking method S4 are as follows. 3.5.1 Embedding algorithm of S4 Input: – – –
–
Gray scale original image Io of size M1 × N1 similar to defined in Section 3.1.1. Watermark (a face image) Wo of size M2 × N2 similar to defined in Section 3.1.1. Embedding locations selection key ckey4 = {ckey4 (m2 , n2 , m3 , n3 ) → (i, j, k); i ∈ {1, 2, · · · M1 } ; j ∈ {1, 2, · · · N1 } ; k ∈ {H, V , D}; m2 = 1, 2, · · · M2 ; n2 = 1, 2, · · · N2 ; m3 = 1, 2, · · · 8; n3 = 1, 2, · · · R(mm3 ) (Section 3.3.1)}, where, Nw = 3 × M1 × N1 . Watermark embedding strength α4 .
Processing: 1. Apply 1-level RDWT on the image Io to obtain Iˆ similar to discussed in Section 3.1.1. 2. We have embedded one bit in each position of each detailed band of the image. Therefore, number ofavailable embedding locations for watermarking is Nw = 3 × M1 × N1 . w Compute d = M2N×N . 2 3. Apply algorithm A1 on Wo to convert it in the binary form W. ˆ 4. Define Iˆw = I. 5. Update Iˆw by embedding W according to formula (9), where (m2 , n2 , m3 , n3 )s and embedding strength are according to Section 3.5.1. 6. Apply 1-level inverse RDWT followed by floating point truncation on the updated Iˆw to obtain the watermarked image Iw . Output: Watermarked image Iw of size M1 × N1 that has same format as of Io . 3.5.2 Extraction algorithm of S4 Input: – – –
An image I of size M1 × N1 similar to defined in Section 3.1.2. Embedding locations selection key ckey4 same as defined in Section 3.5.1. Watermark embedding strength, α4 same as defined in Section 3.5.1.
Multimed Tools Appl
Processing:
1. Apply 1-level RDWT on the image I to obtain Iˆ similar to discussed in Section 3.1.2. 2. Use the formula (10) to extract watermark bits Wex (m2 , n2 , m3 , n3 )s, where (m2 , n2 , m3 , n3 )s and embedding strength are according to Section 3.5.2. 3. Reconstruct watermark We from extracted bits Wex (m2 , n2 , m3 , n3 )s by using weighted binary decoding. Output: Extracted watermark We of size M2 × N2 that has same format as of Wo . 4 Experiments, results and analysis We have performed eight experiments for a detailed analysis of the proposed watermarking methods S1 , S2 , S3 and S4 . We have used peak signal to noise ratio (PSNR) to measure the quality of watermarked images and PSNR and normalized correlation coefficient (NC) to measure the quality of extracted watermarks. Experiment 1 aims to study the effect of watermark embedding strength on the performance of watermarking methods and to find optimum watermark embedding strength for each watermarking method (S1 , S2 , S3 and S4 ). We define optimum watermark embedding strength such that degradation in watermarked image is unnoticeable and distortion in corresponding extracted watermark is minimum (PSNR and NC of extracted watermark are maximum). We have given higher priority to the quality of watermarked image. In Experiment 1, we have also compared the performance of the proposed watermarking methods with existing DWT based watermarking methods. Experiments 2–8 analyse the performance of the proposed watermarking methods against various common attacks/operations such as cropping, Gaussian filtering, Gaussian noise, salt and pepper noise, rotation, JPEG compression and resize respectively. We have performed all the experiments on MATLAB R2007b 64 bit version for windows. In the experiments, we have used a data-set that consists of four host/original images and three watermarks. Figure 1 shows all host images (Im1, im2, im3, Im4) and Fig. 2 shows all watermarks (W1, W2, W3) of the data-set. Each host image is an eight bit gray scale image of size 512 × 512 pixels and each watermark is an eight bit gray scale face image of size 64 × 64 pixels. We have used all the combinations (a total of 12 combinations) of the host images and the watermarks to obtain different watermarked images. The ‘randperm’ function of MATLAB has generated the keys ckey1 , ckey2 , ckey3 and ckey4 . We have used daubechies wavelet filter of length two (db1) for implementing the DWT and RDWT.
Im1 Fig. 1 Original/host images
Im2
Im3
Im4
Multimed Tools Appl
W1
W2
W3
Fig. 2 Original watermarks
4.1 Experiment 1: performance of watermarking methods with watermark embedding strength We have evaluated the performance of S1 and S2 at watermark embedding strengths α1 , α2 = 0.00 : 0.01 : 0.70 and S3 and S4 at watermark embedding strengths α3 , α4 = 0.0 : 0.1 : 4.0. Table 4 gives the quantitative performance of the watermarking methods for the combinations Im1+W1 and Im2+W2 of host image and watermark at various watermark embedding strength. Fig. 3 gives graphical comparison of the performance of the watermarking methods for the combination Im1+W1. The salient observations for each watermarking method are as follows. Watermarking method S1 : PSNR and NC of extracted watermarks increase initially with watermark embedding strength and then become almost constant after watermark embedding strength of 0.3. PSNR of watermarked images decreases with watermark embedding strength. At watermark embedding strength of 0.3, degradation in watermarked image is unnoticeable and distortion in extracted watermark is almost the minimum. Therefore, optimum watermark embedding strength α1 is near α1 ∗ = 0.3. Watermarking method S2 : PSNR and NC of extracted watermarks increase initially, attain maximum values and then decrease with watermark embedding strength. PSNR and NC of extracted watermark attains the maximum value near watermark embedding strength of 0.28. Near α2 = 0.28, degradation in watermarked image is slightly noticeable. PSNR of watermarked images decreases with watermark embedding strength. We have observed that near α2 = 0.05 the degradation in watermarked image is unnoticeable. Therefore, optimum watermark embedding strength α2 is near α2 ∗ = 0.05. Watermarking method S3 : PSNR and NC of extracted watermarks are oscillatory and attain several local and global maxima. Global maxima are attained near α3 = 1, 2, 2.5, 2.6, 3, 3.2, 3.7, 4. PSNR of watermarked images decreases with watermark embedding strength except watermark embedding strength range [0.4 − 0.6]. In the watermark embedding strength (α3 ) range [0.4 − 0.6], the PSNR of watermarked images is slightly oscillatory. Near watermark embedding strength α3 = 1, the degradation in watermarked image is unnoticeable and distortion in extracted watermark is global minimum. Therefore, the optimum watermark embedding strength α3 is near α3 ∗ = 1. Watermarking method S4 : PSNR and NC of extracted watermarks consist of small oscillations and attain several local maxima. The values of local maxima increase with watermark embedding strength. The PSNR of watermarked images decreases with watermark embedding strength except watermark embedding strength range [0.4 − 0.6]. In the watermark embedding strength (α4 ) range [0.4 − 0.6], the PSNR of watermarked images is slightly oscillatory. Near α4 = 2.9, degradation in watermarked image is unnoticeable
Multimed Tools Appl Table 4 Quantitative analysis of the watermarking methods with respect to watermark embedding strength Embedding
PSNR of
PSNR of
NC of
PSNR of
PSNR of
strength
watermarked
extracted
extracted
watermarked
extracted
NC of extracted
image
watermark
watermark
image
watermark
watermark
(dB)
(dB)
(dB)
(dB)
Combination Im1+W1 Watermarking method S1
Watermarking method S2
0.01
56.17
−8.44
0.0780
48.52
15.35
0.9625
0.05
53.80
2.98
0.2928
45.43
29.24
0.9985
0.1
50.06
5.16
0.4690
41.30
35.06
0.9996
0.2
44.89
5.90
0.6131
35.93
41.14
0.9999
0.3
41.53
6.05
0.6589
32.53
43.38
0.9999
0.4
39.10
6.10
0.6773
30.10
39.20
0.9998
0.5
37.17
6.12
0.6865
28.23
33.09
0.9994
0.6
35.63
6.13
0.6914
26.77
28.73
0.9985
0.7
34.30
6.14
0.6944
25.54
25.93
0.9972
Watermarking method S3
Watermarking method S4
0.3
Inf
8.00
0.7746
Inf
7.87
0.7629
0.4
69.06
8.34
0.7936
86.08
7.64
0.7486
0.5
57.14
12.30
0.9236
63.12
9.38
0.8450
0.6
58.00
15.89
0.9662
65.72
8.83
0.8194
1
52.37
Inf
1
55.21
12.34
0.9230
1.5
49.74
40.19
0.9999
52.24
17.79
0.9785
2
47.59
Inf
1
50.28
16.49
0.9707
2.5
45.82
Inf
1
48.60
19.94
0.9868
3
44.33
Inf
1
47.18
18.74
0.9826
3.5
43.05
64.25
1.0000
45.94
19.20
0.9844
4
41.90
Inf
1
44.82
17.26
0.9757
Combination Im2+W2 Watermarking method S1
Watermarking method S2
0.01
55.61
−9.21
0.0102
46.65
15.20
0.9194
0.05
54.48
3.60
0.1505
45.55
29.01
0.9961
0.1
52.23
7.27
0.3000
43.27
35.19
0.9991
0.2
48.06
9.08
0.4864
39.10
40.03
0.9997
0.3
44.98
9.53
0.5773
36.00
38.77
0.9996
0.4
42.65
9.70
0.6229
33.73
36.52
0.9993
0.5
40.77
9.80
0.6484
31.87
34.30
0.9989
0.6
39.27
9.84
0.6630
30.40
32.38
0.9983
0.7
37.96
9.87
0.6729
29.13
30.89
0.9976
Multimed Tools Appl Table 4
(continued)
Embedding
PSNR of
PSNR of
NC of
PSNR of
PSNR of
strength
watermarked
extracted
extracted
watermarked
extracted
extracted
image
watermark
watermark
image
watermark
watermark
(dB)
(dB)
(dB)
(dB)
Watermarking method S3
NC of
Watermarking method S4
0.3
Inf
8.50
0.6347
Inf
8.67
0.6084
0.4
68.51
9.00
0.7373
85.69
8.99
0.6870
0.5
57.22
12.33
0.8417
63.75
9.71
0.7174
0.6
57.71
15.53
0.9157
65.35
9.73
0.7348
1
52.29
Inf
1
55.12
12.42
0.8361
1.5
49.66
41.31
0.9998
52.21
16.42
0.9286
2
47.52
Inf
1
50.18
15.44
0.9075
2.5
45.76
Inf
1
48.51
16.92
0.9351
3
44.23
Inf
1
47.08
16.54
0.9289
3.5
42.93
44.42
0.9999
45.81
16.31
0.9253
4
41.80
Inf
1
44.73
15.01
0.8990
and distortion in extracted watermark is local minimum and almost global minimum. Therefore, optimum watermark embedding strength α4 is near α4 ∗ = 2.9. Table 5 provides quantitative comparison of the proposed watermarking methods for all 12 combinations of host image and watermark. Figures 4 and 5 show watermarked images and corresponding extracted watermarks respectively for two combinations (Im1+W1, Im2+W2). We have observed that watermarked images are not visually degraded (Fig. 4), the watermarks extracted using S1 are very noisy and difficult to recognize (Fig. 5a), the watermarks extracted using S2 are blurred but can easily be recognized (Fig. 5b), the watermarks extracted using S3 are very close to original embedded watermarks (Fig. 5c), and the
a
b S1
c INF
1
90
S3 S4
80 70 60 50 40 30 20
60
S3
S2 S4
40 20
0.9 0.8
S1 S2
0.7
S3
0.6
S4
0.5 0.4
0
10 0
80
S1
NC of extracted watermark
S2
PSNR of extracted watermark
PSNR of watermarked image
INF
0
0.5
1
1.5
2
Embedding strength
2.5
3
0
0.5
1
1.5
2
Embedding strength
2.5
3
0.3 0
0.5
1
1.5
2
2.5
3
Embedding strength
Fig. 3 Performance of watermarking methods (S1 , S2 , S3 and S4 ) at different values of watermark embedding strengths for the combination Im1+W1. a PSNRs of watermarked images. b PSNRs of extracted watermarks. c NCs of extracted watermarks
45.08
44.19
41.51
44.98
44.097
41.57
45.11
44.24
41.22
44.45
43.67
Im1+W3
Im2+W1
Im2+W2
Im2+W3
Im3+W1
Im3+W2
Im3+W3
Im4+W1
Im4+W2
Im4+W3
42.50
42.67
42.01
45.94
46.24
44.86
45.28
45.55
44.37
46.67
47.02
45.43
(α1 = 0.3, α2 = 0.05, α3 = 1, α4 = 2.9)
41.53
Im1+W2
52.37
52.37
52.35
52.10
52.10
51.92
52.29
52.29
52.22
52.39
52.38
52.37
47.55
47.53
47.51
47.38
47.41
47.15
47.44
47.44
47.34
47.57
47.56
47.52
S4
8.12
8.97
5.69
8.64
9.60
5.98
8.60
9.53
5.94
8.73
9.71
6.05
29.24
28.95
29.11
29.08
28.96
28.71
29.27
29.01
29.34
29.36
29.05
29.24
S2
84.25
84.25
Inf
63.15
77.26
78.23
Inf
Inf
Inf
Inf
Inf
Inf
S3
S1
S3
S1
S2
PSNR of extracted watermark (dB)
PSNR of watermarked image
(dB)
Im1+W1
Combination
Table 5 Quantitative performance of the proposed watermarking methods
19.50
19.97
19.13
14.11
14.90
12.15
17.90
18.05
16.67
19.86
19.45
20.20
S4
0.4911
0.4683
0.5664
0.6055
0.5923
0.6442
0.5944
0.5773
0.6366
0.6213
0.6131
0.6589
S1
0.9971
0.9961
0.9984
0.9970
0.9961
0.9983
0.9971
0.9961
0.9985
0.9972
0.9962
0.9984
S2
0.9999
0.9999
1
0.9999
0.9999
0.9999
1
1
1
1
1
1
S3
NC of extracted watermark
0.9734
0.9702
0.9842
0.9002
0.8959
0.9186
0.9603
0.9522
0.9718
0.9776
0.9699
0.9884
S4
Multimed Tools Appl
Multimed Tools Appl
Combination Im1+W1
Combination Im2+W2
a
b
c
d
Fig. 4 Watermarked images. a Watermarking method S1 , α1 = 0.3. b Watermarking method S2 , α2 = 0.05. c Watermarking method S3 , α3 = 1. d Watermarking method S4 , α4 = 2.9
watermarks extracted using S4 are noisy but can be recognized (Fig. 5d). We can rank the performance of watermarking methods as follows: S3 > S2 > S4 > S1 . These results ensure that both DWT based watermarking methods have secured first and second rank respectively, which proves that DWT based watermarking methods can outperform RDWT based watermarking methods. This is a breakthrough as recently Cao et al. [7],
Combination Im1+W1
a
Combination Im2+W2 b c
d
Fig. 5 Extracted watermarks. a Watermarking method S1 , α1 = 0.3. b Watermarking method S2 , α2 = 0.05. c Watermarking method S3 , α3 = 1. d Watermarking method S4 , α4 = 2.9
Multimed Tools Appl Table 6 Comparison of the proposed watermarking methods with Lin et al. [39] and Ma et al. [42] S1
S2
S3
S4
[39]
[42]
Embedding domain
RDWT
DWT
DWT
RDWT
DWT
DWT
Category of
Blind
Blind
Blind
Blind
Blind
Blind
Face
Face
Face
Face
Face
Binary
watermarking method Type of watermark Size of host image
image
image
image
image
image
512 × 512
512 × 512
512 × 512
512 × 512
560 × 296 512 × 512
logo
64 × 64
64 × 64
64 × 64
64 × 64
8×8
(pixels) Size of watermark
32 × 16
(pixels) Size of watermark (bits) PSNR of
64 × 64 × 8 64 × 64 × 8 64 × 64 × 8 64 × 64 × 8 8 × 8 × 8 = 32,768
= 32,768
= 32,768
= 32,768
32 × 16
= 512
= 512
41.53* dB
45.42* dB
52.37* dB
47.52* dB
44.25 dB
48.07 dB
0.6589*
0.9985*
1*
0.9884*
1
1
watermarked image NC of extracted watermark *combination Im1+W1
Parker et al. [49], Vatsa et al. [63], Yang et al. [66] have claimed the superiority of RDWT based watermarking methods over DWT based watermarking methods. For Experiment 1, the performance of the watermarking method S3 is the best. The use of the weighted binary coding has drastically improved the performance of watermarking methods (compare S3 with S2 and S4 with S1 ). Table 6 compares the proposed watermarking methods with the existing DWT based watermarking methods (Lin et al. [39] and Ma et al. [42]). The proposed watermarking method S3 embeds 64 times larger watermark, maintains distortion in extracted watermark at same level and provides better imperceptibility in the watermarked images. The other proposed watermarking methods (S1 , S2 and S4 ) have the advantage of large watermark size and disadvantage of distortion in extracted watermarks. In rest of the paper, we have reported the results for the combination Im1+W1. We have obtained very close results for the other eleven combinations of host image and watermark. 4.2 Experiment 2: cropping from center This experiment performs cropping from center on the watermarked images. In cropping from the center, we have blackened a certain percentage of pixels from center of the watermarked images. We have used cropping percentage ≈ 10 : 10 : 60. Table 7 describes the quantitative performance of the watermarking methods S1 , S2 , S3 , and S4 under cropping for the combination Im1+W1 at different watermark embedding strengths. Better PSNR/NC of extracted watermark ensure better robustness. We observe that robustness of S1 and S2 against cropping improve with watermark embedding strength.
8.029
60.12
8.029
15.33
α1 = 0.25
8.029
60.12
8.025
15.29
α2 = 0.25
8.029
60.12
8.029
15.34
α3 = 2
15.34
8.029
10.01
60.12
α4 = 1
8.029
15.34
α4 = 2
Watermarking method S4
15.34
10.01
α3 = 1
Watermarking method S3
15.33
10.01
α2 = 0.1
Watermarking method S2
15.34
10.01
α1 = 0.1
8.029
15.34
α4 = 3
8.029
15.34
α3 = 3
8.019
15.21
α2 = 0.4
8.028
15.32
α1 = 0.4
percentage
Watermarking method S1
PSNR of watermarked image
(dB)
Cropping
8.029
15.34
α4 = 4
8.029
15.33
α3 = 4
8.006
15.07
α2 = 0.6
8.026
15.3
α1 = 0.6
6.24
11.08
α4 = 1
6.855
15.77
α3 = 1
5.244
9.808
α2 = 0.1
3.9
4.99
α1 = 0.1
(dB)
6.436
13.15
α4 = 2
6.853
15.79
α3 = 2
6.162
13.02
α2 = 0.25
4.657
5.767
α1 = 0.25
6.641
14.15
α4 = 3
6.852
15.73
α3 = 3
6.273
13.59
α2 = 0.4
4.761
5.857
α1 = 0.4
PSNR of extracted watermark
Table 7 Cropping results for the combination Im1+W1 at various watermark embedding strengths
6.466
13.51
α4 = 4
6.852
15.75
α3 = 4
6.301
13.7
α2 = 0.6
4.801
5.888
α1 = 0.6
0.6245
0.8937
α4 = 1
0.6843
0.9653
α3 = 1
0.5346
0.8705
α2 = 0.1
0.186
0.437
α1 = 0.1
0.648
0.9361
α4 = 2
0.6841
0.9654
α3 = 2
0.6142
0.9339
α2 = 0.25
0.348
0.607
α1 = 0.25
NC of extracted watermark
0.6678
0.9491
α4 = 3
0.684
0.9649
α3 = 3
0.6254
0.9419
α2 = 0.4
0.4
0.642
α1 = 0.4
0.6524
0.9417
α4 = 4
0.684
0.9651
α3 = 4
0.6284
0.9435
α2 = 0.6
0.4252
0.6565
α1 = 0.6
Multimed Tools Appl
Multimed Tools Appl
b S1
50
S2
45
S3
40
S4
35 30 25 20 15 10 5 0
10
20
30
40
50
60
c 1
INF
S1 S2 S3
40
S4 30
20
10
0 0
10
20
30
40
50
Cropping percentage
Cropping percentage
60
NC of extracted watermark
55
PSNR of extracted watermark
PSNR of attacked watermarked image
a
0.9 0.8 0.7 0.6 0.5 0.4
S1 S2 S3 S4
0
10
20
30
40
50
60
Cropping percentage
Fig. 6 Performance of watermarking methods (S1 , S2 , S3 and S4 ) for the combination Im1+W1 near respective optimum watermark embedding strength (α1 = 0.3, α2 = 0.05, α3 = 1, α4 = 2.9) under cropping attack on the watermarked images. a PSNRs of cropped watermarked images for different values of cropping percentage. b PSNRs of watermarks extracted from cropped watermarked images for different values of cropping percentage. c NCs of watermarks extracted from cropped watermarked images for different values of cropping percentage
Robustness of S3 is constant at watermark embedding strengths of 1, 2, 3, 4. Robustness of S4 is maximum near watermark embedding strength of 3. Figure 6 gives graphical comparison of the performance of each proposed watermarking method against cropping attack for the combination Im1+W1 near respective optimum watermark embedding strength. The salient observations from Fig. 6 are as follows. Watermarking method S1 : PSNR and NC of extracted watermarks decrease with increase in cropping percentage. S1 has the poorest performance except at cropping percentage of 50 %. At cropping of 50 %, S1 is better than S2 . Watermarking method S2 : PSNR and NC of extracted watermarks have zigzags. S2 under-performs S3 and S4 and it outperforms S1 except at except at cropping percentage of 50 %. Watermarking method S3 : It has the best performance. PSNR and NC of extracted watermarks decrease with increase in cropping percentage. Watermarking method S4 : It has the second best performance. Its performance is lower than the performance of S3 . PSNR and NC of extracted watermarks decrease with increase in cropping percentage. Based on Fig. 6, we can rank the performance of watermarking methods as follows: S3 > S4 > S2 > S1 . 4.3 Experiment 3: Gaussian filtering This experiment applies Gaussian filters of window size 3 × 3 and of different variance on the watermarked images. We have used variance = 0.1 : 0.1 : 1.0. Table 8 describes the performance of watermarking methods S1 , S2 , S3 , and S4 under Gaussian filtering for the combination Im1+W1 at different watermark embedding strengths. We observe that robustness of S1 against Gaussian filtering improves with watermark embedding strength. Robustness of S2 is maximum near watermark embedding
33.38
1
33.36
43.23
α1 = 0.25
39.55
33.33
0.5
1
33.19
36.65
34.14
α2 = 0.25
40.82
33.41
0.5
1
33.4
40.63
47.59
α3 = 2
55.14
48.3
33.41
0.3
0.4
1
α4 = 1
33.41
47.15
50.26
α4 = 2
Watermarking method S4
52.35
0.3
α3 = 1
Watermarking method S3
41.3
0.3
α2 = 0.1
Watermarking method S2
46.85
0.4
α1 = 0.1
33.4
45.83
47.17
α4 = 3
33.4
40.31
44.33
α3 = 3
32.95
33.88
30.28
α2 = 0.4
33.33
40.14
α1 = 0.4
(dB)
variance
Watermarking method S1
PSNR of watermarked image
Filter
33.4
44.58
44.82
α4 = 4
33.39
39.95
41.91
α3 = 4
32.5
31.05
26.93
α2 = 0.6
33.27
37.02
α1 = 0.6
6.519
9.033
12.35
α4 = 1
7.085
8.062
35.48
α3 = 1
4.494
10.73
34.88
α2 = 0.1
3.79
4.96
α1 = 0.1
(dB)
5.48
11.04
16.48
α4 = 2
5.801
8.358
35.44
α3 = 2
4.904
11.09
37.82
α2 = 0.25
4.245
5.7
α1 = 0.25
5.153
13.39
18.74
α4 = 3
5.42
10.91
Inf
α3 = 3
4.945
11.07
34.56
α2 = 0.4
4.298
5.781
α1 = 0.4
PSNR of extracted watermark
Table 8 Gaussian filter results for the combination Im1+W1 at various watermark embedding strengths
4.876
10.72
17.22
α4 = 4
5.192
10.4
Inf
α3 = 4
4.942
10.86
27.54
α2 = 0.6
4.316
5.81
α1 = 0.6
0.6552
0.8292
0.9231
α4 = 1
0.7075
0.7824
0.9996
α3 = 1
0.2906
0.9613
0.9996
α2 = 0.1
0.077
0.426
α1 = 0.1
0.519
0.8926
0.9706
α4 = 2
0.5681
0.7902
0.9996
α3 = 2
0.5808
0.9869
0.9998
α2 = 0.25
0.17
0.622
α1 = 0.25
NC of extracted watermark
0.4609
0.939
0.9826
α4 = 3
0.5112
0.8892
1
α3 = 3
0.7184
0.9898
0.9998
α2 = 0.4
0.252
0.668
α1 = 0.4
0.3998
0.8855
0.9755
α4 = 4
0.4681
0.8745
1
α3 = 4
0.7969
0.9897
0.9985
α2 = 0.6
0.3429
0.6869
α1 = 0.6
Multimed Tools Appl
Multimed Tools Appl
strength of 0.25 up to the filter variance of 0.5. Robustness of S3 improves with watermark embedding strength up to the filter variance of 0.5. Robustness of S4 is maximum near watermark embedding strength of 3 up to the filter variance of 0.4. Figure 7 gives graphical comparison of the performance of each proposed watermarking method against Gaussian filtering for the combination Im1+W1 near respective optimum watermark embedding strength. The salient observations from Fig. 7 are as follows. Watermarking method S1 : Up to the Gaussian filter variance of 0.7, S1 has the poorest performance and after that S1 has better performance than S2 . PSNR and NC of extracted watermarks decrease with increase in Gaussian filter variance. Watermarking method S2 : The comparative performance of S2 is not well behaved. Up to the Gaussian filter variance of 0.3, S2 has the second best performance. In Gaussian filter variance of interval 0.4–0.5, S2 has the best performance. After the variance of 0.5, performance of S2 drops abruptly and after the variance of 0.8, the performance of S2 is the poorest. PSNR and NC of extracted watermarks decrease with increase in Gaussian filter variance. Watermarking method S3 : Up to the Gaussian filter variance of 0.3, the performance of S3 is the best. At variance of 0.4, S2 and S4 have better performance than S3 and at variance of 0.5, the performance of S3 is the second best and lower than the performance of S2 . Beyond the variance of 0.5, the performance of S3 becomes the best again. The PSNR and NC of extracted watermarks decrease with increase in Gaussian filter variance. Watermarking method S4 : Up to the Gaussian filter variance of 0.3, the performance of S4 is lower than the performance of S2 and S3 , and is better than the performance of S1 . At the variance of 0.4, the performance of S4 is the second best and lower than the performance of S2 . In the variance interval of 0.4–0.6, S2 and S3 have better performance. Beyond variance of 0.6, the performance of S4 is the second best and lower than the performance of S3 . PSNR and NC of extracted watermarks decrease with increase in Gaussian filter variance.
b S1 S2 50
S3 S4
45
40
35
30 0
0.2
0.4
0.6
0.8
Gaussian filter variance
1
c INF
1
45
0.9
NC of extracted watermark
55
PSNR of extracted watermark
PSNR of attacked watermarked image
a
S1
40
S2
35
S3
30
S4
25 20 15 10 5 0
0.2
0.4
0.6
0.8
Gaussian filter variance
1
0.8 0.7 0.6 0.5
S1
0.4
S2
0.3
S3 S4
0.2 0.1 0
0.2
0.4
0.6
0.8
1
Gaussian filter variance
Fig. 7 Performance of watermarking methods (S1 , S2 , S3 and S4 ) for the combination Im1+W1 near respective optimum watermark embedding strength (α1 = 0.3, α2 = 0.05, α3 = 1, α4 = 2.9) after Gaussian filtering on watermarked images. a PSNRs of Gaussian filtered watermarked images for different values of Gaussian filter variance. b PSNRs of watermarks extracted from Gaussian filtered watermarked images for different values of Gaussian filter variance. c NCs of watermarks extracted from Gaussian filtered watermarked images for different values of Gaussian filter variance
39.56
1.0
38.2
42.15
α1 = 0.25
37.57
1
33.05
33.93
α2 = 0.25
39.69
1
39.25
45.41
α3 = 2
48.45
39.82
0.1
1
α4 = 1
39.57
46.84
α4 = 2
Watermarking method S4
47.66
0.1
α3 = 1
Watermarking method S3
40.69
0.1
α2 = 0.1
Watermarking method S2
46.75
0.1
α1 = 0.1
39.2
45.19
α4 = 3
38.6
43.18
α3 = 3
29.68
30.05
α2 = 0.4
36.48
38.73
α1 = 0.4
variance ×10−4
Watermarking method S1
PSNR of watermarked image
(dB)
Noise
38.74
43.54
α4 = 4
37.82
41.19
α3 = 4
26.58
26.75
α2 = 0.6
34.26
35.45
α1 = 0.6
8.001
8.185
α4 = 1
8.136
8.548
α3 = 1
19.72
28.33
α2 = 0.1
5.01
5.13
α1 = 0.1
(dB)
7.937
10.35
α4 = 2
7.838
18.21
α3 = 2
27.74
36.49
α2 = 0.25
5.973
5.997
α1 = 0.25
8.38
14.21
α4 = 3
8.644
41.06
α3 = 3
31.31
37.12
α2 = 0.4
6.083
6.097
α1 = 0.4
PSNR of extracted watermark
Table 9 Gaussian noise results for the combination Im1+W1 at various watermark embedding strengths
8.834
15.01
α4 = 4
10.41
63.92
α3 = 4
27.59
28.51
α2 = 0.6
6.129
6.129
α1 = 0.6
0.7793
0.7891
α4 = 1
0.7875
0.8091
α3 = 1
0.9865
0.9981
α2 = 0.1
0.449
0.465
α1 = 0.1
0.7689
0.873
α4 = 2
0.7661
0.9803
α3 = 2
0.9979
0.9997
α2 = 0.25
0.633
0.641
α1 = 0.25
0.8074
0.9505
α4 = 3
0.8164
0.9999
α3 = 3
0.999
0.9998
α2 = 0.4
0.672
0.677
α1 = 0.4
NC of extracted watermark
0.8173
0.9587
α4 = 4
0.8788
1
α3 = 4
0.998
0.9984
α2 = 0.6
0.6905
0.6911
α1 = 0.6
Multimed Tools Appl
Multimed Tools Appl
b 52
S1
50
S2 S3
48
S4
46 44 42 40 38 36 0
c 1
INF
0.2
0.4
0.6
0.8
1
Gaussian noise variancex 10−4
S1
NC of extracted watermark
54
PSNR of extracted watermark
PSNR of attacked watermarked image
a
S2
35
S3 30
S4
25 20 15 10 5 0 0
0.2
0.4
0.6
0.8
1
Gaussian noise variancex 10−4
0.95 0.9 0.85 0.8 0.75 0.7 0.65 0
S1 S2 S3 S4 0.2
0.4
0.6
0.8
1
Gaussian noise variancex 10−4
Fig. 8 Performance of watermarking methods (S1 , S2 , S3 and S4 ) for the combination Im1+W1 near respective optimum watermark embedding strength (α1 = 0.3, α2 = 0.05, α3 = 1, α4 = 2.9) after Gaussian noise addition in watermarked images. a PSNRs of Gaussian noisy watermarked images for different values of Gaussian noise variance. b PSNRs of watermarks extracted from Gaussian noisy watermarked images for different values of Gaussian noise variance. c NCs of watermarks extracted from Gaussian noisy watermarked images for different values of Gaussian noise variance
Based on Fig. 7, we can rank the performance of watermarking methods as follows: S3 > S2 > S4 > S1 when Gaussian filter variance is less than 0.4, S2 > S3 > S4 > S1 when Gaussian filter variance is from 0.4 to 0.5, S3 > S4 > S1 > S2 when Gaussian filter variance is greater than 0.6. 4.4 Experiment 4: Gaussian noise This experiment adds Gaussian noise of zero mean and of different variance in the watermarked images. We have used Gaussian noise variance = 10−5 : 10−5 : 10−4 . Table 9 describes the performance of watermarking methods S1 , S2 , S3 , and S4 under Gaussian noise for the combination Im1+W1 at different watermark embedding strengths. We observe that robustness of S1 against Gaussian noise improves with watermark embedding strength. Robustness of S2 is maximum near watermark embedding strength of 0.4. Robustness of S3 and S4 improves with watermark embedding strength. Figure 8 gives graphical comparison of performance of each proposed watermarking method against Gaussian noise attack on the combination Im1+W1 near respective optimum watermark embedding strength. From Fig. 8, we observe that the PSNR and NC of extracted watermarks decrease with increase in Gaussian noise. Based on Fig. 8, we can rank the performance of watermarking methods as follows. S2 > S4 > S3 > S1 . 4.5 Experiment 5: salt and pepper noise This experiment mixes salt and pepper noise of different density in the watermarked images. We have used salt and pepper noise density = 0.05 : 0.05 : 1.0.
5.662
0.95
5.671
18.54
α1 = 0.25
5.675
0.95
5.671
18.32
α2 = 0.25
15.47
5.664
0.1
0.95
5.674
15.41
18.5
α3 = 2
18.44
5.662
0.05
0.95
α4 = 1
5.664
18.4
α4 = 2
Watermarking method S4
18.49
0.05
α3 = 1
Watermarking method S3
18.42
0.05
α2 = 0.1
Watermarking method S2
18.35
0.05
α1 = 0.1
5.673
18.47
α4 = 3
5.676
15.43
18.44
α3 = 3
5.674
18.24
α2 = 0.4
5.665
18.45
α1 = 0.4
density
Watermarking method S1
PSNR of watermarked image
(dB)
Noise
5.665
18.38
α4 = 4
5.671
15.44
18.47
α3 = 4
5.664
17.91
α2 = 0.6
5.668
18.35
α1 = 0.6
4.13
11.5
α4 = 1
4.314
26.75
42.23
4.121
14.39
α4 = 2
4.137
24.77
35.98
α3 = 2
6.328 −6.09
−1.39 −13.8 α3 = 1
α2 = 0.25
3.183 −6.25
−2.404 −13.81
α2 = 0.1
α1 = 0.25
α1 = 0.1
(dB)
6.965
16.08
α4 = 3
7.791
24.78
38.01
α3 = 3
−2.55
10.12
α2 = 0.4
−2.73
4.649
α1 = 0.4
PSNR of extracted watermark
4.122
15.29
α4 = 4
4.139
26.28
37.78
α3 = 4
−0.11
13.34
α2 = 0.6
−0.19
5.442
α1 = 0.6
Table 10 Salt and pepper noise results for the combination Im1+W1 at various watermark embedding strengths
0.158
0.9063
α4 = 1
0.2626
0.9973
0.9999
α3 = 1
0.0207
0.4653
α2 = 0.1
0.002
0.116
α1 = 0.1
0.1288
0.952
α4 = 2
0.128
0.9957
0.9997
α3 = 2
0.0373
0.7669
α2 = 0.25
−0.01
0.293
α1 = 0.25
NC of extracted watermark
0.8872
0.9677
α4 = 3
0.8536
0.9957
0.9998
α3 = 3
0.023
0.8838
α2 = 0.4
0.013
0.395
α1 = 0.4
0.1326
0.9613
α4 = 4
0.1096
0.997
0.9998
α3 = 4
0.0274
0.9396
α2 = 0.6
0.011
0.514
α1 = 0.6
Multimed Tools Appl
Multimed Tools Appl
b
55 S2
45
S3
40
S4
35 30 25 20 15 10 5 0
S1 S2
40
NC of extracted watermark
50
c INF
S1
PSNR of extracted watermark
PSNR of attacked watermarked image
a
S3 S4
30 20 10 0 −10
1
S1
0.9
S2
0.8
S3
0.7
S4
0.6 0.5 0.4 0.3 0.2 0.1 0
0.2
0.4
0.6
0.8
Salt and pepper noise density
1
−20 0
0.2
0.4
0.6
0.8
Salt and pepper noise density
1
0
0.2
0.4
0.6
0.8
1
Salt and pepper noise density
Fig. 9 Performance of watermarking methods (S1 , S2 , S3 and S4 ) for the combination Im1+W1 near respective optimum watermark embedding strength (α1 = 0.3, α2 = 0.05, α3 = 1, α4 = 2.9) after mixing salt and pepper noise in watermarked images. a PSNRs of noisy watermarked images for different values of salt and pepper noise density. b PSNRs of watermarks extracted from noisy watermarked images for different values of salt and pepper noise density. c NCs of watermarks extracted from noisy watermarked images for different values of salt and pepper noise density
Table 10 describes the performance of watermarking methods S1 , S2 , S3 , and S4 under salt and pepper noise for the combination Im1+W1 at different watermark embedding strengths. We observe that robustness of S1 and S2 against salt and pepper noise improve with watermark embedding strength. Robustness of S3 is almost constant at watermark embedding strength of 1, 2, 3, 4. Robustness of S4 is maximum near watermark embedding strength of 3. Figure 9 gives graphical comparison of performance of each watermarking method against salt and pepper noise attack for the combination Im1+W1 near respective optimum watermark embedding strength. From Fig. 9, we observe that the PSNR and NC of extracted watermarks decrease with increase in salt and pepper noise density. PSNR and NC curves of extracted watermarks for S1 and S2 have small zigzags also. Based on Fig. 9, we can rank the performance of watermarking methods as follows: S3 > S4 > S1 > S2 . 4.6 Experiment 6: rotation This experiment rotates the watermarked images in a counterclockwise direction around their center point and then crop to the size of original image. We have used rotation degree = −10 : 2 : 10. Table 11 describes the performance of watermarking methods S1 , S2 , S3 , and S4 against rotation for the combination Im1+W1 at different watermark embedding strengths. We observe that robustness of S1 and S2 against rotation improves with watermark embedding strength and robustness of S3 and S4 is constant at watermark embedding strength of 1, 2, 3, 4. Figure 10 gives graphical comparison of performance of each watermarking method against rotation for the combination Im1+W1 near respective optimum watermark
17.71
17.95
12.39
−2
2
10
12.39
17.93
17.7
12.27
α1 = 0.25
17.7
17.93
12.39
−2
2
10
12.36
17.85
17.62
12.24
α2 = 0.25
12.27
17.7
17.94
12.39
−10
−2
2
10
α3 = 1
12.39
17.94
17.7
12.27
α3 = 2
Watermarking method S3
12.26
−10
α2 = 0.1
Watermarking method S2
12.27
−10
α1 = 0.1
12.39
17.93
17.7
12.26
α3 = 3
12.32
17.7
17.48
12.2
α2 = 0.4
12.38
17.91
17.68
12.26
α1 = 0.4
(dB)
degree
Watermarking method S1
PSNR of watermarked image
Rotation
12.38
17.92
17.69
12.26
α3 = 4
12.25
17.43
17.22
12.14
α2 = 0.6
12.37
17.88
17.64
12.25
α1 = 0.6
7.777
8.09
8.181
7.75
α3 = 1
2.873
2.675
2.958
2.834
α2 = 0.1
2.17
2.829
3.02
2.368
α1 = 0.1
(dB)
7.573
7.935
8.043
7.502
α3 = 2
3.844
3.757
3.817
3.815
α2 = 0.25
3.749
3.901
3.927
3.745
α1 = 0.25
7.743
8.126
8.103
7.76
α3 = 3
3.963
3.909
3.932
3.944
α2 = 0.4
3.972
4.04
4.042
3.95
α1 = 0.4
PSNR of extracted watermark
Table 11 Rotation results for the combination Im1+W1 at various watermark embedding strengths
7.532
7.912
7.893
7.436
α3 = 4
4.006
3.97
3.978
3.993
α2 = 0.6
4.053
4.088
4.082
4.03
α1 = 0.6
0.7632
0.7877
0.7906
0.7631
α3 = 1
0.7454
0.7737
0.7777
0.7434
α3 = 2
0.0082
−0.012
−0.014 0.0169
−0.007
−0.009
α2 = 0.25
0.0073
0.0181
0.0121
−0.027
α1 = 0.25
−0.007
−0.002
α2 = 0.1
0.0068
0.0155
0.0109
−0.027
α1 = 0.1
NC of extracted watermark
0.7645
0.7912
0.789
0.7662
α3 = 3
0.0007
−0.0097
−0.0063
−0.0142
α2 = 0.4
0.0078
0.0208
0.0128
−0.026
α1 = 0.4
0.7438
0.7697
0.7689
0.7369
α3 = 4
−0.006
−0.007
−0.005
−0.018
α2 = 0.6
0.0084
0.0234
0.013
−0.024
α1 = 0.6
Multimed Tools Appl
17.71
17.94
12.39
−2
2
10
12.39
17.94
17.7
12.27
α4 = 2
12.39
17.94
17.7
12.27
12.39
17.93
17.7
12.26
α4 = 4
−: clockwise direction, no sign: counterclockwise direction
12.27
−10
α4 = 1
α4 = 3
(dB)
Watermarking method S4
PSNR of watermarked image
degree
(continued)
Rotation
Table 11
7.833
8.378
8.358
7.777
α4 = 1
7.526
8.159
8.173
7.575
α4 = 2
7.81
8.426
8.408
7.825
α4 = 3
PSNR of extracted watermark (dB)
7.067
7.865
7.975
7.138
α4 = 4
0.7648
0.7989
0.7987
0.7611
α4 = 1
0.739
0.782
0.784
0.7423
α4 = 2
0.7626
0.8042
0.8043
0.7638
α4 = 3
NC of extracted watermark
0.7022
0.7628
0.77
0.7079
α4 = 4
Multimed Tools Appl
Multimed Tools Appl
b 55 S1
45
S2 S3
40
S4
35 30 25 20 15 −5
0
5
10
1.2 S1
40
NC of extracted watermark
50
10 −10
c INF
PSNR of extracted watermark
PSNR of attacked watermarked Image
a
S2 S3 S4
30
20
10
0
−10 −10
Rotation degree
−5
0
5
Rotation degree
10
1 0.8 S1
0.6
S2 0.4
S3 S4
0.2 0 −0.2 −10
−5
0
5
10
Rotation degree
Fig. 10 Performance of watermarking methods (S1 , S2 , S3 and S4 ) for the combination Im1+W1 near respective optimum watermark embedding strength (α1 = 0.3, α2 = 0.05, α3 = 1, α4 = 2.9) after rotating the watermarked images. a PSNRs of rotated watermarked images for different values of rotation degree. b PSNRs of watermarks extracted from rotated watermarked images for different values of rotation degree. c NCs of watermarks extracted from rotated watermarked images for different values of rotation degree (−: clockwise direction, no sign: counterclockwise direction)
embedding strength. From Fig. 10, we observe that the PSNR and NC of extracted watermarks decrease with increase in rotation degree. Based on Fig. 10, we can rank the performance of watermarking methods as follows: S4 > S3 > S1 > S2 . 4.7 Experiment 7: JPEG compression This experiment compresses the watermarked images using the JPEG compression of different quality. Table 12 describes the performance of watermarking methods S1 , S2 , S3 , and S4 against the JPEG compression for the combination Im1+W1 at different watermark embedding strengths. We observe that robustness of S1 against the JPEG compression slightly improves with watermark embedding strength. Robustness of S2 improves with watermark embedding strength up to the quality of 80. After this quality, robustness of S2 is maximum near the watermark embedding strength of 0.4. Robustness of S3 is almost constant with respect to watermark embedding strength up to the quality of 90. After this quality, robustness of S3 improves with watermark embedding strength. Robustness of S4 is almost constant with watermark embedding strength. Figure 11 gives graphical comparison of performance of each watermarking method against JPEG compression for the combination Im1+W1 near respective optimum watermark embedding strength. From Fig. 11, we observe that the PSNR and NC of extracted watermarks increase with increase in the JPEG compression quality. Based on Fig. 11, we can rank the performance of watermarking methods as follows: S3 > S4 > S1 > S2 when the JPEG compression quality is less than 80, S2 > S4 > S3 > S1 when the JPEG compression quality is more than 90.
30.4
42.94
95
40.29
30.4
24.24
α1 = 0.25
37.6
39.26
80
95
33.54
32.97
24.25
α2 = 0.25
40.62
43.31
90
95
42.46
40.34
24.25
α3 = 2
24.25
40.72
43.55
1
90
95
α4 = 1
43.06
40.56
24.25
α4 = 2
Watermarking method S4
24.25
1
α3 = 1
Watermarking method S3
24.25
1
α2 = 0.1
Watermarking method S2
24.25
10
α1 = 0.1
42.37
40.32
24.25
α4 = 3
41.22
39.86
24.25
α3 = 3
29.87
29.13
24.24
α2 = 0.4
37.75
30.39
24.24
α1 = 0.4
41.41
39.94
24.25
α4 = 4
39.73
39.08
24.25
α3 = 4
26.67
26.12
24.24
α2 = 0.6
34.95
30.37
24.24
α1 = 0.6
8.13
7.601
4.18
α4 = 1
8.073
7.95
4.247
α3 = 1
20.45
6.586
3.827
α2 = 0.1
4.88
3.251
2.915
α1 = 0.1
7.359
6.565
4.21
α4 = 2
7.978
6.948
4.237
α3 = 2
28.65
11.52
4.089
α2 = 0.25
5.963
4.005
3.915
α1 = 0.25
8.523
6.377
4.264
α4 = 3
10.54
7.139
4.294
α3 = 3
32.19
17.57
4.12
α2 = 0.4
6.087
4.11
4.049
α1 = 0.4
(dB)
Watermarking method S1
PSNR of extracted watermark (dB)
PSNR of watermarked image
1
Quality
8.939
6.196
4.224
α4 = 4
13.69
7.624
4.16
α3 = 4
27.85
21.16
4.13
α2 = 0.6
6.128
4.145
4.098
α1 = 0.6
Table 12 JPEG compression results for the combination Im1+W1 at various watermark embedding strengths
0.7841
0.7482
0.1202
α4 = 1
0.7843
0.774
0.173
α3 = 1
0.9883
0.6906
0.0092
α2 = 0.1
0.417
0.7264
0.6569
0.1491
α4 = 2
0.7724
0.6973
0.1667
α3 = 2
0.9982
0.915
0.0222
α2 = 0.25
0.637
0.029
0.006
−0.003 0.014
α1 = 0.25
α1 = 0.1
NC of extracted watermark
0.8014
0.6398
0.1814
α4 = 3
0.8806
0.7119
0.2006
α3 = 3
0.9992
0.9789
0.0346
α2 = 0.4
0.676
0.052
0.013
α1 = 0.4
0.8191
0.6165
0.157
α4 = 4
0.9432
0.7449
0.1039
α3 = 4
0.9981
0.991
0.0472
α2 = 0.6
0.691
0.076
0.021
α1 = 0.6
Multimed Tools Appl
Multimed Tools Appl
b
1
40 38 36 34
S1
32
S2 S3
30
S4
28 26 20
40
60
80
Quality
100
S1 12
NC of extracted watermark
42
24 0
c 14
44
PSNR of extracted watermark
PSNR of attacked watermarked image
a
S2 S3 S4
10
8
6
4
2 0
20
40
60
80
100
Quality
0.9 0.8 0.7
S1 S2 S3 S4
0.6 0.5 0.4 0.3 0.2 0.1 0 0
20
40
60
80
100
Quality
Fig. 11 Performance of watermarking methods (S1 , S2 , S3 and S4 ) for the combination Im1+W1 near respective optimum watermark embedding strength (α1 = 0.3, α2 = 0.05, α3 = 1, α4 = 2.9) after JPEG compression of watermarked images. a PSNRs of JPEG compressed watermarked images for different values of compression Quality. b PSNRs of watermarks extracted from JPEG compressed watermarked images for different values of compression Quality. c NCs of watermarks extracted from JPEG compressed watermarked images for different values of compression Quality
4.8 Experiment 8: resize This experiment first scales down the watermarked images and then scales up to the original size. We have used scale = 0.1 : 0.1 : 1.0. Table 13 describes the performance of watermarking methods S1 , S2 , S3 , and S4 against resize operation for the combination Im1+W1 at different watermark embedding strengths. We observe that robustness of S1 against resize operation slightly improves with watermark embedding strength. Robustness of S2 is almost maximum near watermark embedding strength of 0.4. Robustness of S3 and S4 is maximum near watermark embedding strength of 1. Figure 12 gives graphical comparison of performance of each watermarking method against resize operation for the combination Im1+W1 near respective optimum watermark embedding strength. From Fig. 12, we observe that the PSNR and NC of extracted watermarks increase with increase in resize scale (except resize scale near 0.5). Based on Fig. 12, we can rank the performance of watermarking methods as follows: S3 > S4 > S2 > S1 . 4.9 Comparison of PSNR and NC of extracted watermarks From the Figs. 3, 6–12, we have observed the following relation between PSNR and NC of extracted watermarks. 1. PSNR of extracted watermarks increases with increase in NC and vice-versa. 2. For the high value of NC of extracted watermarks (near 0.99) the PSNR curves of extracted watermarks have better discriminative capability for performance comparison. 3. For the low value of NC, NC curves of extracted watermarks have better discriminative capability for performance comparison.
39.18
0.9
38.78
20.31
α1 = 0.25
38.51
0.9
36.35
20.31
α2 = 0.25
39.25
0.9
39.12
20.31
α3 = 2
20.31
39.29
0.1
0.9
α4 = 1
39.23
20.32
α4 = 2
Watermarking method S4
20.31
0.1
α3 = 1
Watermarking method S3
20.31
0.1
α2 = 0.1
Watermarking method S2
20.32
0.1
α1 = 0.1
39.12
20.31
α4 = 3
38.92
20.31
α3 = 3
34
20.31
α2 = 0.4
38.14
20.31
α1 = 0.4
density
Watermarking method S1
PSNR of watermarked image
(dB)
Noise
38.99
20.31
α4 = 4
38.66
20.31
α3 = 4
31.42
20.32
α2 = 0.6
37.05
20.31
α1 = 0.6
7.972
5.488
α4 = 1
8.14
5.61
α3 = 1
7.771
4.07
α2 = 0.1
4.223
3.95
α1 = 0.1
(dB)
6.559
5.107
α4 = 2
7.732
5.229
α3 = 2
8.187
4.119
α2 = 0.25
4.873
4.092
α1 = 0.25
6.265
4.888
α4 = 3
8.028
4.974
α3 = 3
8.206
4.123
α2 = 0.4
4.948
4.109
α1 = 0.4
PSNR of extracted watermark
Table 13 Resize results for the combination Im1+W1 at various watermark embedding strengths
5.794
4.751
α4 = 4
7.811
4.844
α3 = 4
8.116
4.123
α2 = 0.6
4.973
4.115
α1 = 0.6
0.7741
0.5234
α4 = 1
0.7837
0.549
α3 = 1
0.8127
0.026
α2 = 0.1
0.6559
0.4511
α4 = 2
0.7529
0.477
α3 = 2
0.8867
0.0263
α2 = 0.25
0.457
−0
−0.002 0.25
α1 = 0.25
α1 = 0.1
NC of extracted watermark
0.6248
0.4026
α4 = 3
0.7708
0.4247
α3 = 3
0.8965
0.0264
α2 = 0.4
0.544
−0
α1 = 0.4
0.5659
0.3678
α4 = 4
0.7569
0.3927
α3 = 4
0.8987
0.0268
α2 = 0.6
0.591
−0
α1 = 0.6
Multimed Tools Appl
Multimed Tools Appl
b 55 S2
1.2
S3
45
S4
40 35 30 25
0.2
0.6
0.4
0.8
1
S1
S1
45
NC of extracted watermark
50
20
c INF
S1
PSNR of extracted watermark
PSNR of attacked watermarked image
a
S2
40
S3
35
S4
30 25 20 15 10 5 0
0.2
0.6
0.4
0.8
S2
1
S3 0.8
S4
0.6 0.4 0.2 0 −0.2
1
0.2
0.4
0.6
0.8
1
Scale
Scale
Scale
Fig. 12 Performance of watermarking methods (S1 , S2 , S3 and S4 ) for the combination Im1+W1 near respective optimum watermark embedding strength (α1 = 0.3, α2 = 0.05, α3 = 1, α4 = 2.9) after resizing the watermarked images. a PSNRs of resized watermarked images for different values of resize scale. b PSNRs of watermarks extracted from resized watermarked images for different values of resize scale. c NCs of watermarks extracted from resized watermarked images for different values of resize scale
Table 14 Performance ranking of watermarking methods
Attack
S1
S2
S3
S4
First rank
No attack
4
2
1
3
S3 (DWT)
Cropping
4
3
1
2
S3 (DWT)
Gaussian filter
4
2
1
3
S3 (DWT)
4
1
2
3
S2 (DWT)
3
4
1
2
S3 (DWT)
Gaussian noise
4
1
3
2
S2 (DWT)
Salt and pepper noise
3
4
1
2
S3 (DWT)
Rotation
3
4
2
1
S4 (RDWT)
JPEG compression
3
4
1
2
S3 (DWT)
4
1
3
2
S2 (DWT)
4
3
1
2
S3 (DWT)
(variance < 0.4) Gaussian filter (variance 0.4 to 0.5) Gaussian filter (variance > 0.6)
(quality < 80) JPEG compression (quality > 90) Resize
5 Conclusion We have proposed and compared four blind watermarking methods. Eight bit gray scale face image is used as watermark. The size of watermark is 64 × 64 pixels (64 × 64 × 8 bits), which is 64 times larger than the size of watermark in [39, 42]. We have found optimum watermark embedding strength for each proposed watermarking method. The optimum watermark embedding strength for S1 , S2 , S3 and S4 is 0.3, 0.05, 1 and 2.9 respectively. Without attack, watermarking method S3 has the outstanding performance; PSNR value of watermarked image is more than 50 dB and PSNR and NC value of extracted watermark is INF dB and 1 respectively.
Multimed Tools Appl
The performance ranking of each watermarking method under different scenarios (without attack, cropping, Gaussian filtering, Gaussian noise, salt and pepper noise, rotation, JPEG compression and resize) is given in Table 14. One breakthrough is that DWT based watermarking methods have dominated except rotation attack. The weighted binary coding has drastically improved the performance of watermarking methods compared with watermarking methods of same transform domain. This signifies the very importance of weighted binary coding. PSNR and NC of extracted watermarks increase and decrease together. For higher value of NC, PSNR curves provide better distinctiveness for performance comparison of watermarking methods and vice-versa. For some attacks, method S2 has obtained highest robustness and for other attacks (except rotation), method S3 has obtained highest robustness. In future, we can try to develop more robust watermarking method using the advantages of the watermarking methods S2 and S3 . Moreover, efficient hardware implementation of these watermarking methods for real time watermarking [33, 57] is another important direction.
Acknowledgment One of the authors, Himanshu Agarwal, acknowledges the University Grants Commission (UGC), of New Delhi, India for granting him a scholarship under the JRF scheme for his research.
References 1. Unique Identification Authority of India Planning Commission Government of India (2009) Unique identification authority of India (UIDAI). http://uidai.gov.in/. Accessed 1 Jan 2014 2. Bao P, Ma X (2005) Image adaptive watermarking using wavelet domain singular value decomposition. Trans Circuits Syst Video Technol 15(1):96–102 3. Barni M, Bartolini F, Piva A (2001) Improved wavelet-based watermarking through pixel-wise masking. IEEE Trans Image Process 10(5):783–791 4. Bhatnagar G, Raman B (2009) A new robust reference watermarking scheme based on dwt-svd. Comput Stand Interfaces 31(5):1002–1013 5. Bhatnagar G, Raman B (2011) A new robust reference logo watermarking scheme. Multimed Tools Appl 52(2–3):621–640 6. Bhatnagar G, Wu Q, Raman B (2012) A new robust adjustable logo watermarking scheme. Comput Secur 31(1):40–58 7. Cao JG, Fowler JE, Younan NH (2001) An image-adaptive watermark based on a redundant wavelet transform. In: Proceedings of IEEE international conference on image processing 2001, vol 2. Thessaloniki, pp 277–280 8. Cheddad A, Condell J, Curran K, McKevitt P (2008) Biometric inspired digital image steganography. In: Proceedings of IEEE 15th annual international conference and workshop on the Engineering of Computer Based Systems, ECBS 2008. Belfast, pp 159–168 9. Chung Y, Moon D, Moon K, Pan S (2005) Hiding biometric data for secure transmission. In: Knowledgebased intelligent information and engineering systems. Lecture notes in computer science, vol 3683, pp 1049–1057 10. Dawei Z, Guanrong C, Wenbo L (2004) A chaos-based robust wavelet-domain watermarking algorithm. Chaos, Solitons Fractals 22(1):47–54 11. Ding S, Li C, Liu Z (2010) Protecting hidden transmission of biometrics using authentication watermarking. In: Proceeding of WASE International Conference on Information Engineering (ICIE) 2010, vol 2. Beidaihe, pp 105–108 12. Dugad R, Ratakonda K, Ahuja N (1998) A new wavelet-based scheme for watermarking images. In: Proceedings of IEEE international conference on image processing, 1998, vol 2. Chicago, pp 419– 423 13. Fallahpour M, Meg´ıas D (2011) High capacity audio watermarking using the high frequency band of the wavelet domain. Multimed Tools Appl 52(2–3):485–498
Multimed Tools Appl 14. Fouad M, El Saddik A, Petriu E (2010) Combining dwt and lsb watermarking to secure revocable iris templates. In: Proceddings of IEEE 10th international conference on Information Sciences Signal Processing and their Applications (ISSPA), 2010. Kuala Lumpur, pp 25–28 15. Fridrich J (2000) Visual hash for oblivious watermarking. In: Wong PW, Delp EJ (eds) SPIE international society for optics and photonics, security and watermarking of multimedia contents II, vol 3971. San Jose, pp 286–294 16. Ghouti L, Bouridane A, Ibrahim MK, Boussakta S (2006) Digital image watermarking using balanced multiwavelets. IEEE Trans Signal Process 54(4):1519–1536 17. Gunsel B, Uludag U, Murat Tekalp A (2002) Robust watermarking of fingerprint images. Pattern Recogn 35(12):2739–2747 18. Hassanien AE (2006) Hiding iris data for authentication of digital images using wavelet theory. Pattern Recogn Image Anal 16(4):637–643 19. Holliman M, Memon N (2000) Counterfeiting attacks on oblivious block-wise independent invisible watermarking schemes. IEEE Trans Image Process 9(3):432–441 20. Hong I, Kim I, Han SS (2001) A blind watermarking technique using wavelet transform. In: IEEE International Symposium on Industrial Electronics (ISIE), 2001, vol 3. Pusan, pp 1946–1950 21. Hsieh MS, Tseng DC, Huang YH (2001) Hiding digital watermarks using multiresolution wavelet transform. IEEE Trans Ind Electron 48(5):875–882 22. Hsu CT, Wu JL (1998) Multiresolution watermarking for digital images. IEEE Trans Circ Syst II Analog Digit Signal Process 45(8):1097–1101 23. Jain AK, Nandakumar K (2012) Biometric authentication: system security and user privacy. IEEE Comput 45(11):87–92 24. Jain AK, Uludag U (2003) Hiding biometric data. IEEE Trans Pattern Anal Mach Intell 25(11):1494– 1498 25. Jain AK, Ross A, Prabhakar S (2004) An introduction to biometric recognition. IEEE Trans Circuits Syst Video Technol 14(1):4–20 26. Keyvanpour MR, Merrikh-Bayat F (2011) Robust dynamic block-based image watermarking in dwt domain. Procedia Comput Sci 3:238–242 27. Khan MK, Xie L, Zhang J (2007) Robust hiding of fingerprint-biometric data into audio signals. In: Advances in biometrics lecture notes in computer science, vol 4642, pp 702–712 28. Khan MK, Zhang J, Tian L (2007) Chaotic secure content-based hidden transmission of biometric templates. Chaos Solitons Fractals 32(5):1749–1759 29. Kim JR, Moon YS (1999) A robust wavelet-based digital watermarking using level-adaptive thresholding. In: Proceedings of IEEE International Conference on Image Processing (ICIP), 1999, vol 2. Kobe, pp 226–230 30. Kim T, Chung Y, Jung S, Moon D (2006) Secure remote fingerprint verification using dual watermarks. In: Digital rights management technologies, issues, challenges and systems. Lecture notes in computer science, vol 3919. Springer, pp 217–227 31. Kim Wg, Lee H (2009) Multimodal biometric image watermarking using two-stage integrity verification. Signal Process 89(12):2385–2399 32. Korus P, Białas J, Dziech A (2012) A new approach to high-capacity annotation watermarking based on digital fountain codes. Multimed Tools Appl 1–19. doi:10.1007/s11042-011-0986-8 33. Kougianos E, Mohanty SP, Mahapatra RN (2009) Hardware assisted watermarking for multimedia. Comput Electr Eng 35(2):339–358 34. Kundur D, Hatzinakos D (1999) Digital watermarking for telltale tamper proofing and authentication. Proc IEEE 87(7):1167–1180 35. Kundur D, Hatzinakos D (2004) Toward robust logo watermarking using multiresolution image fusion principles. IEEE Trans Multimedia 6(1):185–198 36. Lai CC, Tsai CC (2010) Digital image watermarking using discrete wavelet transform and singular value decomposition. IEEE Trans Instrum Meas 59(11):3060–3063 37. Li C, Wang Y, Liu L (2009) A biometric templates secure transmission method based on bi-layer watermarking and pki. In: IEEE international conference on multimedia information networking and security, 2009, MINES’09, vol 2. Hubei, pp 95–98 38. Li CT, Yang FM, Lee CS (2002) Oblivious fragile watermarking scheme for image authentication. In: IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2002, vol 4. Orlando, pp IV–3445–IV–3448 39. Lin WH, Horng SJ, Kao TW, Fan P, Lee CL, Pan Y (2008) An efficient watermarking method based on significant difference of wavelet coefficient quantization. IEEE Trans Multimedia 10(5):746–757 40. Lin WH, Wang YR, Horng SJ, Kao TW, Pan Y (2009) A blind watermarking method using maximum wavelet coefficient quantization. Expert Syst Appl 36(9):11509–11516
Multimed Tools Appl 41. Liu JL, Lou DC, Chang MC, Tso HK (2006) A robust watermarking scheme using self-reference image. Comput Stand Interfaces 28(3):356–367 42. Ma B, Wang Y, Li C, Zhang Z, Huang D (2013) Secure multimodal biometric authentication with wavelet quantization based fingerprint watermarking. Multimed Tools Appl 1–30. doi:10.1007/s11042-013-1372-5 43. Memon N, Wong PW (2001) A buyer-seller watermarking protocol. IEEE Trans Image Process 10(4):643–649 44. Mohanty SP (2009) A secure digital camera architecture for integrated real-time digital rights management. J Syst Archit 55(10):468–480 45. Mohanty SP, Bhargava BK (2008) Invisible watermarking based on creation and robust insertionextraction of image adaptive watermarks. ACM Trans. Multimed Comput Commun Appl 5(2):12 46. Morimoto N (1999) Digital watermarking technology with practical applications. Informing Sci Spec Issue Multimed Informing Technol Part 1 2(4):107–111 47. Noore A, Singh R, Vatsa M, Houck MM (2007) Enhancing security of fingerprints through contextual biometric watermarking. Forensic Sci Int 169(2):188–194 48. Park KR, Jeong DS, Kang BJ, Lee EC (2007) A study on iris feature watermarking on face data, vol 4432 49. Parker KM, Fowler JE (2005) Redundant-wavelet watermarking with pixel-wise masking. In: IEEE International Conference on Image Processing, ICIP 2005, vol 2005, pp I685–I688 50. Podilchuk CI, Zeng W (1998) Image-adaptive watermarking using visual models. IEEE J Select Areas Commun 16(4):525–539 51. Prabhakar S, Pankanti S, Jain AK (2003) Biometric recognition: security and privacy concerns. IEEE Secur Priv 1(2):33–42 52. Qi M, Lu Y, Du N, Zhang Y, Wang C, Kong J (2010) A novel image hiding approach based on correlation analysis for secure multimodal biometrics. J Netw Comput Appl 33(3):247–257 53. Rani A, Raman B, Kumar S (2013) A robust watermarking scheme exploiting balanced neural tree for rightful ownership protection. Multimed Tools Appl 1–24. doi:10.1007/s11042-013-1528-3 54. Ratha NK, Connell JH, Bolle RM (2000) Secure data hiding in wavelet compressed fingerprint images. In: Proceedings of the 2000 ACM workshops on multimedia. ACM, USA, pp 127–130 55. Ratha NK, Connell JH, Bolle RM (2001) Enhancing security and privacy in biometrics-based authentication systems. IBM Syst J 40(3):614–634 56. Reddy AA, Chatterji BN (2005) A new wavelet based logo-watermarking scheme. Pattern Recogn Lett 26(7):1019–1027 57. Roy SD, Li X, Shoshan Y, Fish A, Yadid-Pecht O (2013) Hardware implementation of a digital watermarking system for video authentication. IEEE Trans Circ Syst Video Technol 22(2):289–301 58. Shejul AA, Kulkarni UL (2010) A dwt based approach for steganography using biometrics. In: IEEE international conference on Data Storage and Data Engineering (DSDE) 2010. Bangalore, pp 39–43 59. Soheili MR (2008) Redundant watermarking using wavelet packets. In: IEEE/ACS International Conference on Computer Systems and Applications, AICCSA. Doha, pp 591–598 60. Subramanyam A, Emmanuel S, Kankanhalli MS (2012) Robust watermarking of compressed and encrypted jpeg2000 images. IEEE Trans Multimed 14(3):703–716 61. Tay R, Havlicek J (2002) Image watermarking using wavelets. In: IEEE the 45th Midwest Symposium on Circuits and Systems, MWSCAS 2002, vol 3, pp III–258–III–261 62. Vatsa M, Singh R, Mitra P, Noore A (2004) Comparing robustness of watermarking algorithms on biometrics data. In: Workshop proceedings biometrics: challenges arising from theory to practice. Cambridge, pp 5–8 63. Vatsa M, Singh R, Noore A (2009) Feature based rdwt watermarking for multimodal biometric system. Image Vis Comput 27(3):293–304 64. Wang N, Men C (2013) Reversible fragile watermarking for locating tampered blocks in 2d vector maps. Multimed Tools Appl 67(3):709–739. doi:10.1007/s11042-012-1333-4 65. Wei Z, Qin P, Fu Y (1998) Perceptual digital watermark of images using wavelet transform. IEEE Trans Consumer Electron 44(4):1267–1272 66. Yang Hy, Wang Xy, Wang Cp (2012) A robust digital watermarking algorithm in undecimated discrete wavelet transform domain. Comput Electr Eng 39(3):893–906 67. Zebbiche K, Ghouti L, Khelifi F, Bouridane A (2006) Protecting fingerprint data using watermarking. In: First NASA/ESA conference on Adaptive Hardware and Systems, AHS 2006. Istanbul, pp 451– 456 68. Zebbiche K, Khelifi F, Bouridane A (2008) An efficient watermarking technique for the protection of fingerprint images. EURASIP J Inf Secur 4:1–20. doi:10.1155/2008/918601 69. Zhang XD, Feng J, Lo KT (2003) Image watermarking using tree-based spatial-frequency feature of wavelet transform. J Vis Commun Image Represent 14(4):474–491
Multimed Tools Appl
Himanshu Agarwal is a full time Ph.D. student in the Department of Mathematics, Indian Institute of Technology Roorkee since July 2009. He is a member of Computer Vision Graphics and Image Processing Laboratory in the Department of Mathematics, Indian Institute of Technology Roorkee since July 2009. He was a visiting research student in the Applied Computer Science Department, The University of Winnipeg, Canada in 2012-2013 for six months. He has received M.Sc degree in Industrial Mathematics & Informatics from Indian Institute of Technology Roorkee in 2009. His areas of research include digital watermarking, biometrics and applications of wavelets in images.
Balasubramanian Raman is an Associate Professor in the Department of Computer Science and Engineering at Indian Institute of Technology Roorkee, India. He has obtained MSc degree in Mathematics from Madras Christian College (University of Madras) in 1996 and PhD from Indian Institute of Technology Madras in 2001. He was a post doctoral fellow at University of Missouri Columbia, USA in 2001–2002 and a post doctoral associate at Rutgers, the State University of New Jersey, USA in 2002–2003. He joined Department of Mathematics at Indian Institute of Technology Roorkee as lecturer in 2004 and became assistant professor in 2006 and associate professor in 2012. He was a visiting professor and a member of Computer Vision and Sensing Systems Laboratory in the Department of Electrical and Computer Engineering at University of Windsor, Canada during May–August 2009. So far he has published more than 150 research papers in reputed Journals and Conferences. His area of research includes Computer Vision- Optical Flow Problems, Fractional Transform Theory, Wavelet Analysis, Image and Video Processing, Multimedia Security: Digital Image Watermarking and Encryption, Biometrics, Content Based Image and Video Retrieval, Hyperspectral & Microwave Imaging, Visualization and Volume Graphics.
Multimed Tools Appl
Ibrahim Venkat (formerly K. Venkatasubramanian) received the Ph. D. degree from Heriot-Watt University, Edinburgh, UK in 2010. Since 2010, he is serving as a Senior Academic at the School of Computer Sciences, Universiti Sains Malaysia. His research interests include bioinspired computing, biometrics, computer vision, and pattern recognition.