Neural network based robust image watermarking technique in LWT domain

Abstract

In this paper, a robust image watermarking technique has been proposed in lifting wavelet transform (LWT) domain. Neural network is incorporated in the watermark extraction process to achieve improved robustness against different attacks. The integration of neural network with LWT makes the system robust to various attacks maintaining an adequate level of imperceptibility. The 3-level LWT coefficients are randomized and arranged in 2×2 non-overlapping blocks. Each block is modified according to a binary watermark bit. Randomization of coefficients and blocks has been done to enhance the security of the system. The binary watermark bit is also encrypted using another key. The scheme provides an average imperceptibility of 43.88 dB for a watermark capacity of 512 bits. The robustness has been observed against all the intentional and non-intentional attacks. The technique provides satisfactory robustness against different attacks such as noising attacks, de-noising attacks, lossy compression attacks, image processing attacks and some geometric attacks. The algorithm has been tested on a large image database containing different class of images.

Keywords

Lifting wavelet transform (LWT)image watermarking artificial neural network (ANN)peak signal to noise ratio (PSNR)normalized cross-correlation (NC)

1 Introduction

The advancement of computer technologies and internet over the past few decades have changed the world and daily lives significantly by making the capture, storage and distribution of multimedia data very easy and convenient. The easy access of multimedia data brings with itself the challenge of many practical issues like copyright protection and content authentication. Digital watermarking [6] is a useful technique to fulfill the demand of copyright protection. In watermarking, we embed some information into digital data which provides ownership claim and authentication to prevent it from illicit and unauthorized copying.

Watermarking technique can be divided into various categories depending on the type of classification. Depending on the type of extraction algorithm used, a watermarking scheme can be a blind or a semi-blind or a non-blind technique. The original image is not required in blind algorithm for watermark extraction. The watermarking technique can also be classified into robust, fragile and semi-fragile techniques. A robust watermarking technique should have the capability to survive both the non-intentional and intentional attacks. The watermark embedding can be carried out either in spatial domain or transform domain. Spatial domain techniques are easy to implement. In contrast, transform domain techniques are complex but provide better robustness.

In recent times, numerous transform domain techniques [3 , 16] based on discrete Fourier transform (DFT), lifting wavelet transform (LWT), discrete wavelet transform (DWT), discrete cosine transform (DCT) have been proposed. These techniques utilize human perception property and signal characteristics so as to obtain enhanced imperceptibility and robustness. Different decomposition techniques [2, 13] like QR decomposition, singular value decomposition (SVD) etc. also have been integrated with transformation technique to achieve enhanced performance.

DWT has drawn a lot of attention in designing robust watermarking systems due to its multi-resolution properties in time and frequency. On the other hand, DCT has good energy compaction property. The combination of both the DWT and DCT for watermarking generally exhibits good performance in terms of both imperceptibility and robustness. However, it increases computational complexity. The transformation technique LWT has both multi-resolution as well as energy compaction property. So, in our proposed work, LWT has been used as a watermark embedding domain.

Over the past few years, there has been a tendency of using soft computing and machine learning approaches to achieve superior performances. The machine learning based approaches have been applied in different fields of image processing applications like hand gesture recognition, digital watermarking, image restoration, etc. [1 , 17]. In recent period, watermarking techniques based on different machine learning approaches have been proposed. Using multi-layer perceptron (MLPs) based neural networks, a color image watermarking technique has been proposed by Yu et al. [12]. In this paper, MLPs based ANN is used to learn the relation between embedded watermark bit and watermarked image. Using Genetic algorithm (GA) and DWT, Ramanjaneyulu et al. [8] have proposed a scheme for watermark embedding which provides improved performances over the previous methods. In this method, GA is used to optimize different scaling factors used in embedding and extraction processes. Though, the technique provides improved robustness against various types of image processing operations, but, fails to provide sufficient robustness against JPEG attack and image sharpening operation. Huang et al. [15] have proposed a novel blind watermarking technique using Back Propagation neural network in wavelet domain. In this paper, a scrambled watermark is embedded using the advantage of Human Vision System (HVS) to achieve better imperceptibility and robustness. Neural network is used to memorize the relation between the embedded watermark and corresponding watermarked image. Different SVM based robust image watermarking techniques [11, 17] have also been proposed in the literature. Peng et al. [4] have proposed a novel image watermarking technique in multi-wavelet domain based on SVM. The algorithm have utilized special frequency band and the property of image for watermarking. Though the scheme is reasonably robust against various attacks but fails to achieve robustness against average filtering, median filtering, JPEG attacks and scaling attack effectively. Yang et al. [5] have also proposed a robust technique in undecimated discrete wavelet transform (UDWT) domain using fuzzy SVM for geometric distortion correction. Though the technique provides adequate robustness, yet it requires excessive computational time and also it is not robust to local geometric distortions.

In this paper, we propose an image watermarking technique robust to both intentional and non-intentional attacks. The main objective of this paper is to design a system with enhanced features of robustness and to extract the watermark with maximum possible correlation. LWT coefficients are modified in such a way so that it has less effect on the invisibility of the watermarking system. In this scheme, watermark extraction has been considered as a binary classification problem. The advantage of the scheme is that it uses computationally simple feature for training and testing the ANN. Instead of conventional DWT, LWT has been used in the proposed approach as it is fast and requires less memory. The main motivation of using ANN is due to its good generalization capability and generalization performance even when the watermarked image is vigorously distorted. Different simple statistical features like mean, variance, entropy etc. of a LWT block have been used as an input feature for training and testing of ANN. Both the training and testing of ANN are performed on contaminated environment i.e. on the attacked watermarked image to obtain improved robustness against different attacks. The contribution of the proposed scheme is as follows.

Energy compaction property of LWT and good generalization ability of ANN together can resist more image distortions. The integration of ANN for watermark extraction with LWT helps to achieve robustness against different attacks that maintains adequate level of imperceptibility.

The security of the proposed technique has been enhanced by randomizing the LWT coefficients and blocks using different keys. This randomization is necessary for making it difficult for impostor to find the locations of watermark even though the imposter have enough knowledge of watermark embedding and extraction process.

The proposed technique is oblivious in nature. During watermark extraction, neither original host nor the watermark is required.

The proposed technique has been tested on an image database of 300 images collected from CVG-UGR image database, USC-SIPI image database, and standard database available at www.imageprocessingplace.com. The performance of the algorithm has been observed on different class of images.

Rest of the paper is organized as follows. The detailed watermark embedding and extraction process has been discussed in section 2. Section 3 describes the performance of the proposed algorithm. In Section 4, the conclusion and future scope of the proposed techniques have been discussed.

2 Proposed watermarking algorithm

In this paper, a LWT based watermarking scheme incorporating ANN for digital watermark extraction has been proposed. Firstly, LWT has been performed on the original image to decompose it up to the third level. Then, HL3 sub-band has been chosen for watermark embedding purpose as shown in Fig. 1. The detail embedding process is provided in watermark embedding section. During extraction process, binary watermark bits are detected based on ANN classifier. The blocks and coefficients of HL3 sub-band of the watermarked image are found out in a similar way as in embedding process. The training and testing data are prepared by concatenating the feature vectors that provide best possible results along with corresponding block coefficients.

Fig.1

LWT sub band.

2.1 Watermark embedding procedure

In this work, watermark (W) is composed of two components namely the reference information (RW) and the signature information (SW). The reference watermark (RW) bits are generated randomly and it also acts as a key 3. The signature watermark (SW) bits are encrypted by performing ‘XOR’ operation between the one dimensional vector of SW (i.e. logo watermark image) and key 3. The RW and SW have been concatenated and represented as a single unit as in Equation (1):

$\begin{matrix} W & = & RW + SW \\ = & W_{1}; . . . . ., W_{L_{r}}; W_{L_{r} + 1}, . . . . ., W_{L_{r} + L_{s}} \\ = & W_{1}, . . . . ., W_{L_{w}} \end{matrix}$ (1)

In Equation (1), L_w, L_r, L_s are the length of W, RW and SW respectively.

The mathematical formulation for embedding binary watermark bits is shown in Table 1. The largest coefficient of the corresponding block is quantized for each watermark bit.

Table 1

Modification of the largest coefficients

If the watermark bit is ‘1’	If the watermark
bit is ‘0’
c_i (n)	c_i (n)
$= {\begin{matrix} c_{i} (n) + T, & if d_{i max} < max (γ, T) \\ c_{i} (n), & otherwise \end{matrix}$	= c_i (n) - d_imax

In Table 1, T is called as embedding threshold used in quantizing block coefficients, d_imax is the difference between the two largest coefficients present in ith block and c_i (n) is the largest coefficient corresponding to ith block. The average coefficient difference between the two largest coefficients γ is expressed as in Equation (2): $γ = [\frac{\sum_{i = 1}^{L_{W}} d_{i max}}{L_{W}}]$ (2)

In Equation (2), L_w represents the total number of blocks present in HL3 sub-band in which the watermark bits have been embedded. The block diagram of the proposed embedding procedure is demonstrated in Fig. 2. The step by step process of binary watermark embedding has been discussed below:

Fig.2

Watermark embedding procedure.

Step 1. The HL3 sub-band has been obtained by performing LWT three times on the original host image.

Step 2. The coefficients of the 3-level HL3 sub-band are then randomized by using secret seed key 1.

Step 3. The randomized coefficients have been arranged in non-overlapping 2×2 blocks, and all these blocks are again randomized using secret seed key 2.

Step 4. Using Equation (2), the average difference between the two largest coefficients γ is calculated.

Step 5. The reference watermark (RW) and signature watermark (SW) are concatenated and is represented as a one dimensional vector of binary watermark W having length L_w.

Step 6. For all L_w bits of watermark, the following operation is performed:

In each block, the two largest coefficients are considered for modification.

The largest coefficient is modified according to Table 1 depending on the embedded watermark bit i.e. either 0 or 1.

Step 7. All the blocks and updated coefficients of HL3 sub-band are inversely shuffled using the same keys as mentioned in Step 2 and Step 3 .

Step 8. Finally, watermarked image is obtained by performing inverse LWT operation.

2.2 Watermark extraction procedure

The detection of binary watermark bits have been considered as a binary classification problem and the ANN classifier is used for this purpose. The binary watermark bit detection can be realized as an ability of generalization from machine learning viewpoint. ANN is used as a watermark detector in our proposed scheme as it has good generalization ability. So, utilizing ANN may be helpful to attain enhanced robustness of the watermarking system under diverse attack conditions.

The proposed scheme is blind in nature because neither the original image nor the watermark (SW) is required in case of watermark detection. As discussed previously, the two parts of the watermark, i.e. signature information and reference information are embedded into the host image. At first, the blocks of HL3 sub-band in which RW was embedded have been extracted. Then, different statistical parameters such as: mean (p1), standard deviation (p2), variance (p3), covariance (p4), kurtosis (p5), skewness (p6), median (p7), moment (5th order) (p8), quantile (p9), and Energy (p10) of the corresponding blocks are evaluated. The set of parameters (p1; p2;...; p10) is called as features vector and represented as (f₁;...; f₁₀) i.e. feature vector set. The feature vector set is given as input feature to train the ANN machine. The HL3 sub-band contains a total of 1024 blocks in which each block having a size of 2×2. The first 512 blocks are utilized to prepare the training pattern whereas, the next 512 blocks are used to generate the testing pattern. The testing pattern provides extracted signature watermark (SW’). So, the size of feature vector set {p(k)} is 512×10. The training pattern set is created using the feature set along with the coefficients of corresponding block. Similarly, the set of testing pattern is also generated. The training of ANN is performed by applying set of training patterns as an input to the machine and keeping RW bit as a target vector. At last, the trained ANN has been used to classify a set of testing patterns. Using the results provided by the trained ANN, the signature watermark (SW) can be detected.

The block diagram of the proposed extraction scheme is demonstrated in Fig. 3. The steps for watermark extraction is discussed below.

Fig.3

Watermark extraction procedure.

Step 1. HL3 sub-band coefficients are obtained in a similar way as in embedding process by performing 3-level LWT on the watermarked image.

Step 2. The HL3 sub-band coefficients and corresponding blocks are randomized using secret keys (key 1 and key 2).

Step 3. Features set {p_i (k) |k = 1, 2, . . .10 } is generated from the blocks where reference watermark (RW) are embedded.

Step 4. In this step, system is trained using ANN classifier. The feature set generated from the blocks in which RW was embedded has been used as input feature vector. The corresponding bits of RW are used as target vector. The detail of training process is discussed next.

Step 5. The features vectors have been calculated from the blocks in which SW was embedded. Then, by using the well trained ANN corresponding output bits i.e. extracted watermark bits are obtained. The detail testing process is discussed in subsequent paragraph.

In this proposed work, first the system is trained with feature set using ANN comprising of previously described features. The LWT blocks in which RW was embedded has been used for SVM training purpose. The nonlinear behavior present in the pattern can be modeled using the ANN [7] with hidden layers. An ANN is a computational tool inspired by the biological nervous systems to process information. A feed forward network having n layers can be modeled with y_2,t as the neurons in the output layer; I_n being the neurons in the input layer; w_1,ij is the weight matrix between input and hidden layers; B_1,j is the bias values of hidden neurons; w_2,tj is the matrix between the hidden and the output layers, and B_2,t is the bias values of output neurons, then $y_{NET 1} = \sum_{i = 0}^{n} w_{1, ij} I_{i} + B_{1, j}$ (3) $y_{1, j} = f_{j} [y_{NET 1, j}], j = 1, 2, . . ., m$ (4) $y_{NET 2, t} = \sum_{i = 0}^{m} w_{2, tj} y_{1, i} + B_{2, t}$ (5) $y_{2, t} = f_{t} [y_{NET 2, t}], t = 1, 2, . . ., k$ (6)

During training phase, all the 512 blocks in which RW is embedded are assigned to two classes, ‘0’ i.e. 0 bit embedded and ‘1’ i.e. 1 bit embedded. The graphical representation for generating ANN training and testing patterns is shown in Fig. 4.

Fig.4

Block diagram for generating ANN training and testing patterns.

The testing pattern is generated by calculating feature vectors from the blocks in which SW was embedded. The feature vectors along with corresponding blocks coefficients are used for ANN classification. If the classified output is greater than or equal to 0.5, the output is considered as 1 i.e. bit 1 was embedded, otherwise, the output is considered as 0 i.e. bit 0 was embedded. The encrypted extracted watermark bits SW’ are decrypted using key 3 to obtain extracted watermark.

3 Results and discussions

The LWT-ANN based watermarking technique is tested on 200 gray images of size (512×512). Different classes of images such as standard images like Lena, Mandril, Peppers, satellite images, medical images, texture images etc. have been considered in this analysis. These images are collected from CVG-UGR image database, USC-SIPI image database, and www.imageprocessingplace.com. The binary watermarks are collected from the MPEG7_CE_shape descriptor database. All the analysis and experimentation have been carried out in Windows 7 based Matlab (R2013a). The hardware used for simulation purpose is a Dell computer with Intel core i5 processor and 16 GB RAM. For the sake of conciseness, the original Lena image (512×512), watermarked Lena image along with original watermark and extracted watermark under no attack condition is shown in the Fig. 5.

Fig.5

(a) Original image “Lena” of size (512×512), (b) Watermarked image with PSNR = 44.49 dB, (c) Original signature watermark (16×32), and (d) extracted watermark.

Imperceptibility is measured in terms of peak signal to noise ratio (PSNR) between the original cover image and watermarked image. PSNR is measured according to [1]. Robustness is measured in terms of normalized cross correlation (NC) and bit error rate (BER) as calculated in [17] between original watermark and extracted watermark.

3.1 Imperceptibility analysis

The imperceptibility of the algorithm is tested on various image database. It has been observed experimentally that the T (embedding threshold) balances the two performance parameter i.e. imperceptibility and robustness. For lower value of T, imperceptibility increases but robustness decreases, and vice-versa. It has been found that T = 35 gives better balance between imperceptibility and robustness. So, for this value of T, all experiments have been performed. Figure 6 shows the PSNR for 15 different standard images. It is seen from the Fig. 6 that the technique provides adequate imperceptibility irrespective of the type of the images. The performance has been evaluated on 200 images and the average PSNR is found to be 43.88 dB.

Fig.6

The imperceptibility for different images.

3.2 Robustness analysis

Robustness of the proposed scheme has been observed against various attacks. The attacks include both non-intentional and intentional attack like speckle noise (SN), salt and pepper noise (SPN), Gaussian noise (GN), joint photographic experts group (JPEG) with different quality factor, average filtering (AF), Gaussian filtering (GF), histogram equalization (HE), image sharpening (IS), Gamma Correction (GC), scaling (SCL), and cropping (CR) attacks. The robustness test is performed on different types of images. The representative results are shown for 10 different standard images in Tables 2 and 3. From the experimental results, it has been observed that the technique provides superior performance against image processing attacks and JPEG attack. The technique also provides satisfactory performance against noising attacks, de-noising attacks and some kind of geometric attacks. However, the scheme provides slightly degraded robustness against some of the attacks like average filtering and scaling attack with 0.5 scaling factor. The robustness test has been carried out on 200 images and the average NC is demonstrated in Tables 4 and 5. Tables 4 and 5 also show the extracted watermark SW’ on Lena watermarked image against different attacks.

Table 2
NC values of extracted watermark for different attacks over 10 images

Images JPEG 20 JPEG 30 JPEG 40 JPEG 50 AF (3×3) GF (3×3) SPN (0.01) SN (0.01)

Lena 0.9686 0.9765 0.9804 0.9882 0.8678 0.9921 0.9131 0.9210

Peppers 0.9110 0.9690 0.9764 0.9921 0.9375 0.9608 0.8587 0.9136

Mandril 0.9110 0.9084 0.9419 0.9764 0.6817 0.9842 0.8814 0.9052

Jetplane 0.9566 0.9725 0.9684 0.9961 0.8854 0.9961 0.8276 0.8387

Barbara 0.8022 0.8362 0.9076 0.9261 0.7824 0.9498 0.7986 0.8893

Lake 0.9373 0.9566 0.9725 0.9766 0.8501 0.9842 0.9019 0.8935

X-ray 0.8884 0.9212 0.9686 0.9803 0.9289 0.9803 0.8054 0.8396

Galaxy 0.9012 0.9142 0.9317 0.9502 0.8815 0.9843 0.8429 0.9253

Livingroom 0.8987 0.9052 0.9330 0.9690 0.8229 0.9805 0.8695 0.8499

Elaine 0.8893 0.9803 0.9803 0.9728 0.9369 1.0000 0.8814 0.9012

Images	JPEG 20	JPEG 30	JPEG 40	JPEG 50	AF (3×3)	GF (3×3)	SPN (0.01)	SN (0.01)
Lena	0.9686	0.9765	0.9804	0.9882	0.8678	0.9921	0.9131	0.9210
Peppers	0.9110	0.9690	0.9764	0.9921	0.9375	0.9608	0.8587	0.9136
Mandril	0.9110	0.9084	0.9419	0.9764	0.6817	0.9842	0.8814	0.9052
Jetplane	0.9566	0.9725	0.9684	0.9961	0.8854	0.9961	0.8276	0.8387
Barbara	0.8022	0.8362	0.9076	0.9261	0.7824	0.9498	0.7986	0.8893
Lake	0.9373	0.9566	0.9725	0.9766	0.8501	0.9842	0.9019	0.8935
X-ray	0.8884	0.9212	0.9686	0.9803	0.9289	0.9803	0.8054	0.8396
Galaxy	0.9012	0.9142	0.9317	0.9502	0.8815	0.9843	0.8429	0.9253
Livingroom	0.8987	0.9052	0.9330	0.9690	0.8229	0.9805	0.8695	0.8499
Elaine	0.8893	0.9803	0.9803	0.9728	0.9369	1.0000	0.8814	0.9012

Table 3

NC values of extracted watermark for different attacks over 10 images

Images	HE	IS	GC	SCL (0.75)	SCL (0.5)	CR (10%)	CR (25%)	CR (50%)
Lena	0.9765	0.9961	0.9726	0.9491	0.5854	0.9245	0.7797	0.5008
Peppers	0.9802	0.9611	0.9961	0.8930	0.7714	0.8541	0.6623	0.4893
Mandril	0.9460	0.9178	0.9843	0.8529	0.4634	0.9145	0.6909	0.5102
Jetplane	0.8715	0.9299	0.9961	0.9765	0.6594	0.8780	0.4579	0.4081
Barbara	0.8447	0.8604	0.9245	0.8952	0.3673	0.8114	0.5181	0.4754
Lake	0.9882	0.9842	0.9921	0.9614	0.5768	0.8507	0.7059	0.5763
X-ray	0.8447	0.9251	0.9842	0.9724	0.7756	0.8262	0.7314	0.6032
Galaxy	0.8786	0.9493	0.9961	0.9725	0.5720	0.9138	0.6832	0.6196
Livingroom	0.9052	0.9685	0.9921	0.9842	0.5181	0.8952	0.7654	0.5531
Elaine	0.9464	0.9345	0.9493	0.9764	0.7566	0.8380	0.8082	0.4483

Table 4

Average robustness (NC) for different attacks and extracted watermark on Lena image

Attacks	JPEG	JPEG	JPEG	JPEG	AF	GF	SPN	SN
	20	30	40	50	(3×3)	(3×3)	(0.01)	(0.01)
NC (average)	0.8992	0.9285	0.9463	0.9678	0.8522	0.9806	0.8517	0.8786
Extracted watermark

Table 5

Average robustness (NC) for different attacks and extracted watermark on Lena image

Attacks	HE	IS	GC	SCL	SCL	CR	CR	CR
				(0.75)	(0.5)	(10%)	(25%)	(50%)
NC (average)	0.9131	0.9386	0.9709	0.9297	0.5933	0.8654	0.6713	0.5109
Extracted watermark

3.3 Comparative performance analysis

For comparative performance analysis, different

recent watermarking techniques based on LWT-SVM, LWT-QR, LWT-SVD etc. have been considered for this purpose. The performance of the proposed technique has been compared with some of the state-of-art watermarking techniques [3 , 18]. The imperceptibility of the proposed scheme has been compared in terms of PSNR (dB) for four standard images which are shown in Fig. 7. Figure 7 suggests that the proposed technique provides superior imperceptibility than the techniques proposed by Kasana et al. [3], Verma et al. [17], Wang et al. [18] and Hamghalam et al. [10] in most of the cases. The robustness against different types of attacks has also been compared with that of the methods [11, 13, 17] for different standard images. Figure 8 shows the robustness comparison in terms of NC with [17] for Lena watermarked image. The robustness is also compared in terms of BER with [11, 13] for 3 different standard images in Table 6. The watermarking technique [11] is the authors’ previous work proposed in LWT-SVD domain. The different types of attacks such as noising attacks, de-noising attacks, lossy compression attacks, image processing attacks and some geometric attacks have been considered during this analysis. The proposed technique provides improved performance against most of the attacks such as JPEG attacks, image processing attacks. The algorithm provides comparable robustness against some attacks like average filtering, and salt and pepper noise. So, the simulation results show that proposed technique gives better solution for proving ownership claim under diverse attack conditions.

Fig.7

Comparison of Imperceptibility with different methods.

Fig.8

Comparison of robustness with Verma et al. [17].

Table 6

The comparison results with [11, 13] for different attacks in terms of BER

Attacks	Lena			Peppers			Mandril
	[13]	[11]	Proposed	[13]	[11]	Proposed	[13]	[11]	Proposed
SPN (0.005)	0.0566	0.0562	0.0253	0.0557	0.0473	0.0351	0.0518	0.0472	0.0390
SPN (0.01)	0.1152	0.1250	0.0429	0.1230	0.1219	0.0703	0.1025	0.0840	0.0625
SPN (0.02)	0.1638	0.1582	0.0859	0.1973	0.1855	0.1250	0.1768	0.1758	0.1445
GN (0.001)	0.0547	0.0467	0.0351	0.0557	0.0464	0.0390	0.0684	0.0572	0.0429
GN (0.005)	0.2490	0.2132	0.2089	0.2637	0.1926	0.1933	0.2480	0.2289	0.2226
GN (0.01)	0.3584	0.3456	0.3476	0.3635	0.3652	0.3535	0.3281	0.2985	0.3105
IS	0.0107	0.0059	0.0020	0.0107	0.0062	0.0136	0.0557	0.0092	0.0410
AF (3×3)	0.0039	0.1660	0.0664	0.0078	0.1699	0.0391	0.0410	0.3379	0.1601
GF (3×3)	0	0.0098	0.0039	0.0020	0.0214	0.0195	0.0315	0.0195	0.0153
JPEG 50	0.0020	0.0020	0.0020	0.0020	0.0020	0.0020	0.0256	0.0117	0.0059
JPEG 70	0	0	0	0	0	0	0.0049	0	0

4 Conclusion

In this paper, a robust blind image watermarking technique using neural network for watermark extraction in LWT domain has been proposed. The 3-level diagonal sub band of LWT transformed image has been chosen for watermark embedding purpose. During watermark extraction, ANN classifier has been incorporated in LWT domain for the classification technique to extract binary watermark with maximum possible correlation. The proposed scheme provides an average imperceptibility of 43.88 dB on a large image database with an embedding capacity of 512 bits. It has been observed that the algorithm provides robustness against different intentional and non-intentional attacks. Experimental results suggest that the proposed technique gives satisfactory performance on different types of images under various signal processing attacks. The performance has also been compared with different state of the art methods. It is seen that the technique provides better performance than other related methods. As a scope of future work, different geometric distortion correction approach may be integrated with the present system to improve robustness against de-synchronization attacks like rotation and translation.

Footnotes

Acknowledgments

The authors would like to acknowledge Speech and Image Processing Laboratory, Department of Electronics and Communication Engineering, National Institute of Technology Silchar, India for providing support and necessary facilities for carrying out this work. Moreover, the authors would like to thank Mr. Mohammad Azharuddin Laskar of the same department for his valuable suggestions.

References

Roy

and Laskar

R.H.

, Non-casual linear prediction based adaptive filter for removal of high density impulse noise from color images, AEU-International Journal of Electronics and Communications72 (2017), 114–124.

Lai

C.C.

and Tsai

C.C.

, Digital image watermarking using discrete wavelet transform and singular value decomposition, IEEE Transactions on Instrumentation and Measurement59(11) (2010), 3060–3063.

Kasana

and Kasana

S.S.

, Reference based Semi Blind Image Watermarking Scheme in Wavelet Domain, Optik142 (2017), 191–204.

Peng

, Wang

and Wang

, Image watermarking method in multiwavelet domain based on support vector machines, Journal of Systems and Software83(8) (2010), 1470–1477.

Yang

H.Y.

, Wang

X.Y.

and Wang

C.P.

, A robust digital watermarking algorithm in undecimated discrete wavelet transform domain, Computers and Electrical Engineering39(3) (2013), 893–906.

Cox

I.J.

, Miller

M.L.

, Bloom

J.A.

and Honsinger

, San Francisco: Morgan Kaufmann, Digital watermarking (Vol. 1558607145), 2002.

Singha

and Laskar

R.H.

, Self co-articulation detection and trajectory guided recognition for dynamic hand gestures, IET Computer Vision10(2) (2016), 143–152.

Ramanjaneyulu

and Rajarajeswari

, Wavelet-based oblivious image watermarking scheme using genetic algorithm, IET Image Processing6(4) (2012), 364–373.

Fattahi

and Latif

, A novel scheme for selection of watermark strength in digital image watermarking based on harmony search algorithm, Journal of Intelligent and Fuzzy Systems28(5) (2015), 2357–2367.

10.

Hamghalam

, Mirzakuchaki

and Akhaee

M.A.

, Geometric modelling of the wavelet coefficients for image watermarking using optimum detector, IET Image Processing8(3) (2014), 162–172.

11.

Islam

and Laskar

R.H.

, Geometric distortion correction based robust watermarking scheme in LWT-SVD domain with digital watermark extraction using SVM, Multimedia Tools and Applications (2017), 1–28. 10.1007/s11042-017-5035-9

12.

P.T.

, Tsai

H.H.

and Lin

J.S.

, Digital watermarking based on neural networks for color images, Signal processing81(3) (2001), 663–671.

13.

Mehta

, Rajpal

and Vishwakarma

V.P.

, LWT-QR decomposition based robust and efficient image watermarking scheme using Lagrangian SVR, Multimedia Tools and Applications75(7) (2016), 4129–4150.

14.

Lin

S.D.

and Chen

C.F.

, A robust DCT-based watermarking for copyright protection, IEEE Transactions on Consumer Electronics46(3) (2000), 415–421.

15.

Huang

, Zhang

, Feng

and Yang

, Blind watermarking scheme based on neural network, Proceedings of the 7th IEEE World Congress on Intelligent Control and Automation (2008), 5985–5989.

16.

Tsui

T.K.

, Zhang

X.P.

and Androutsos

, Color image watermarking using multidimensional Fourier transforms, IEEE Transactions on Information Forensics and security3(1) (2008), 16–28.

17.

Verma

V.S.

, Jha

R.K.

and Ojha

, Digital watermark extraction using support vector machine with principal component analysis based feature reduction, Journal of Visual Communication and Image Representation31 (2015), 75–85.

18.

Wang

X.Y.

, Liu

Y.N.

, Xu

, Wang

A.L.

and Yang

H.Y.

, Blind optimum detector for robust image watermarking in non-subsampled shearlet domain, Information Sciences372 (2016), 634–654.