1 Introduction

1.1 Motivation

The advent of computer networks and the Internet has made digital media production and distribution (e.g. audio, video, image, etc.) quite easy these days [16, 17, 33]. Because of this, authentication and copyright protection of digital media has became a prime concern. To address this issue, digital watermarking can be used [1, 10, 20]. Digital watermarking is the procedure by which the signal (watermark) is embedded into the cover signal. Digital watermarking include two main phases: watermark embedding and watermark extraction [15, 16]. The watermarking scheme can be divided into two classes: spatial domain and frequency domain[17, 34, 36]. Various transform domain-based techniques, such as Discrete Wavelet Transform (DWT), Singular Value Decomposition (SVD), Discrete Cosine Transform (DCT), and Lifting Wavelet Transform (LWT) etc. are used. In comparison with traditional wavelet transform LWT is faster and more effective [13, 34]. The frequency domain is more robust but spatial domain schemes are simpler to implement [6, 23]. One of the key issues about digital image watermarking techniques is the security of a watermark. The appropriate encryption of the watermark is thus necessary in the watermarking system [21, 26].

There is always a trade-off between the features of watermarking [17]. Robustness, imperceptibility security, and complexity are the four major features of the watermarking scheme. Watermarking scheme must have to satisfy two opposing properties of robustness and transparency. To balance these two properties selection of block to embed watermark embedding plays a vital role [25]. Alpha blending techniques can also be used to address this trade-off. [9]. It can be used to incorporate invisible watermarks into the image’s salient characteristics. Alpha blending scheme is resilient against the attacks and provides high protection [2, 4, 5]. Frequency domain methods are helpful in locating the watermark embedding region which ensures the watermarking scheme is more robust. The traditional wavelet transform has high computational and memory requirements using convolution-based implementation. Lifting wavelet transform has been designed to address these disadvantages. LWT provides strong efficiency in digital watermarking as opposed to conventional wavelet transform [34]. To ensure robustness against the number of attacks, Y CbCr color space is used for watermarking. Y CbCr channel is also desirable for watermarking, as it closely models the HVS [27].

1.2 Literature review

Various works have been done in the watermarking to date. In recent years, researchers have studied to improve the color image watermarking scheme. In this section, some of them are briefly discussed.

Kang et al. [12] has proposed a secure and robust color image watermarking scheme. Here, to create the zero watermark, the proposed scheme uses Frobenius norm in the DWT-SVD domain and majority voting. Experimental results shows that this scheme is not robust against the geometric attacks. Pandey et al. [26] proposed a stable hybrid lossless, reliable color watermarking technique lifting method and GWO. In addition, scrambling of the watermark was also achieved by using Arnold transform to improve security and robustness. Robustness of this scheme is comparatively very poor in comparison with other color image watermarking schemes.

For authentication using LWT and SVD, Kejgir et al. [13] has proposed a color image watermarking. Using LWT transform, the cover image is decomposed, and the watermark is inserted into LL subband. Moreover, false positive problem is its main drawback as embedding is carried out in singular component. Liu et al. [20] present a blind color image watermarking based on Schur’s decomposition. In this scheme, color watermark is used for the embedding while affine transform was used to encrypt the watermark. Moreover, both imperceptibility and robustness of this scheme is low.

Wang et al. [38] based on the Fuzzy Least Square Support Vector Machine (FLS-SVM) and Bessel K (BKF) distribution have proposed a robust color image watermarking method. Bessel K’s shape parameter and scale parameter is used as a function vector to train the model. In [30] a hybrid robust watermarking scheme based on is proposed. Here, watermarking is carried out in Y component of Y CbCr color space.

Vaidya et al. [36] suggested a robust and semi-blind watermarking method for color images. Arnold Transform was used to encrypt the watermark author. Although, this scheme is robust against different attacks but gives poor imperceptibility. Koley [14] suggests watermarking of color images in the LWT domain using α −β blending. The technique infuses the watermark with the factors ‘α’ and ‘β’ according to the PC feature map by changing the diagonal information coefficients of the cover image. Experimental results shows that this techniques is not robust against JPEG compression attack.

Chowdhury et al. [3] proposes a blind symmetric watermarking approach in both canonical and cepstrum domains. Using the four-connected t-o’clock process, the security fiber mark is scrambled. Roy et al. [31] have suggested a non-blind, hybrid watermarking technique based on DWT and SVD. Here, component Y is chosen for embedding the watermark. In [28] using Walsh Hadamard Transform (WHT), a color image watermarking scheme is presented. To encrypt watermark 2D-Logistic Sine Coupling Map encryption scheme is used.

Hu et al. [11] developed an SVD-based watermarking scheme which implements mixed modulation to ensure successful extraction of watermarks. For watermarking the color watermark is used here. In [29] DCT based color image watermarking technique is proposed. Here, binary watermark is inserted in Blue and green component of RGB color space. In this scheme computational cost is high. A blind hybrid domain watermarking scheme for color image is proposed in [35], however, the security of this scheme is needed to analyze.

Wang et al. [39] propose a novel blind color image watermarking scheme in DCT domain based on JND. In contrast masking effect, the diversity of orientation between the various blocks is further considered. Laur et al. [19] proposed a watermarking scheme using entropy for imperceptible and robust color image watermarking. Chirp z-transform, orthogonal-triangular decomposition, and SVD are used to embed a watermark in color image DWT. In [8] DFT based color image watermarking is proposed. Here, random sequence are used as a watermark. From its results it can been observed that this techniques is relatively less robust than other frequency domain scheme like DWT, DCT, SVD etc.

Over the past few decades, watermarking methods has come out as authentic method for authentication of data and copyright protection. The literature of watermarking reveals that there is always a trade-off between the various features of watermarking. It is very challenging for research to address these trade-off. Also, to develop a watermarking scheme which is robust against all attacks is very challenging . Wavelet based watermarking technique is comparatively more robust than other frequency domain watermarking techniques. Moreover, in comparison with traditional wavelet transform LWT is computationally efficient. In this work, watermarking is carried out in the Y component of Y CbCr color space. The proposed method involves LWT, entropy, ACM, and alpha blending techniques.

1.3 Contributions and organization

The main objective of this paper is to develop a watermarking scheme which can address the trade-off between the various features of watermarking.

The main contributions to this work can be summarized as follows:

  1. 1)

    A robust and secure color image watermarking using LWT in Y CbCr color space is proposed in this work.

  2. 2)

    Block is selected adaptively for watermark insertion in the Y component of the cover image using the visual entropy and the edge entropy.

  3. 3)

    Alpha blending technique is used to balance the trade-offs between imperceptibility and robustness of the watermarking techniques.

  4. 4)

    For better robustness and less computation time LWT domain is used for watermarking, whereas Arnold’s Cat Map is used to enhance the security of watermarking technique.

The rest of the paper is structured as follows. The preliminaries of LWT, ACM, and Image entropy are discussed in Section 2. Section 3 describes the proposed scheme and examines the experimental result in Section 4. The conclusion is drawn up in Section 5.

2 Background

2.1 Lifting wavelet transform

W. Sweldens proposed an advancement of DWT called Lifting Wavelet Transform. In the lifting wavelet transformation, sampling up and down is simply replaced by splitting and merge into each of the stages [13]. LWT split down the images in four sub-bands [14]. The forward and the reverse LWT are shown in Figs. 1 and 2 correspondingly. In general, the lifting scheme consists of three steps: Divide, Dual lifting, and Primal lift [32].

Fig. 1
figure 1

Block diagram of forward lifting

Full size image
Fig. 2
figure 2

Block diagram of reverse lifting

Full size image

Divide: Here the original signal A(i, j) is split into two samples: even sample Ae(i,j) and odd samples Ao(i,j) and is represented by (1) and (2) respectively.

$$ A_{e}(i,j) = A(i,2j) $$
(1)
$$ A_{o}(i,j) = A(i,2j+1) $$
(2)

Dual Lifting (Predict): We can view this step as a high-pass filtering process. Here, even samples are used to predict odd samples, and the abstract difference is produced as follows:

$$ a(i,j) = A_{o}(i,j) - P[A_{e}(i,j) ] $$
(3)

where P[.] is the predict operator.

Primal Lifting (Update): TThis wavelet lifting step can be viewed as a low-pass filtering procedure. The al(i,j) low frequency part is a coarse approximation to the original signal A(i, j), and this is done by applying U[.] as follows:

$$ a_{l}(i,j) = A_{e}(i,j) - U[a(i,j) ]\\ $$
(4)

2.2 Arnold cat map

ACM also known as Arnold Transform is a form of chaotic encryption which can give watermark security [36]. This encryption technique named after Vladimir I. Arnold. Substantially ACM shears image by a factor of 1 in both vertical and horizontal directions. Further, folds the image collected back to itself [14]. Since ACM retains the size of the watermark even after encryption, this means that the encryption process does not affect the ability of the payload. ACM can be denoted as:

$$ \begin{bmatrix} x_{n+1} \\ y_{n+1} \end{bmatrix} =\begin{bmatrix} 1 & 1 \\ 1 & 2 \end{bmatrix} \begin{bmatrix} x_{n} \\ y_{n} \end{bmatrix}\mod N\\ $$
(5)

where N is order of image matrix and (xn,yn) and (xn+ 1,yn+ 1) represents the pixel value before and after scrambling. Figure 3 illustrate the original watermark and the encrypted watermark image for various iterations of ACM.

Fig. 3
figure 3

a Original grayscale watermark b Encrypted watermark for 10 iterations c Encrypted watermark for 25 iterations d Encrypted watermark for 50 iterations e Encrypted watermark for 75 iterations f Encrypted watermark for 100 iterations

Full size image

2.3 Image entropy

The concept of entropy can be extended to two classes: Visual entropy and edge entropy [7, 22, 23]. Visual entropy and edge entropy consider human visual characteristics. Therefore, it offers the most important host image region for embedding a watermark [6]. Entropy is a proper indicator of adjacent pixel spatial correlation. An image’s visual entropy is described as

$$ E_{v} = -\sum\limits_{i=1}^{L-1}p_{i}\log p_{i} $$
(6)

Edge entropy is an exponential type of entropy capable of holding the image in two dimensional spatial correlations. The edge entropy gives more information about pixels dispersion and edges of the images. Edge entropy or average edge information can be defined as

$$ E_{e} = \sum\limits_{i=1}^{L-1}p_{i} e^{1-p_{i}} $$
(7)

Here ‘pi’ represents the occurrence probability of event i and ‘1-pi’ represents the uncertainty of the pixel value i.

3 Proposed techniques

In this section, an improved color watermarking techniques in the LWT domain is presented. Here, block block is selected using visual entropy and edge entropy to embed watermark. The block having the optimal entropy value BE is selected to embed the watermark. The optimal entropy value is calculated using (8). The proposed technique is classified into the following two stages: watermark embedding and watermark extraction. Figures 4 and 5 shows the overview of the watermark embedding and the watermark extraction process respectively. In the embedding process, ACM image scrambling is used to improve robustness and security. Whereas, its inverse transformation is used in the process of extraction. In the proposed scheme we have taken 50 iterations to encrypt the watermark.

$$ B_{E} = max (E_{v} - E_{e}) $$
(8)
Fig. 4
figure 4

Watermark embedding process

Full size image
Fig. 5
figure 5

Watermark extraction process

Full size image

3.1 Watermark embedding

In this work, watermark embedding is carried out in the Y component of YCbCr color space using LWT. Here, cover image is transformed into YCbCr color space from RGB color space. Y component is selected and first level of LWT is applied over it. Y component is selected because in comparison of Cb and Cr, inserting watermark in Y components yields better robustness. Further, Y component is divided into non overlapping block of size 32×32. To embed watermark, block in HH component of cover image (Y component) is selected adaptively using visual entropy and edge entropy. A grayscale watermark of size 64×64 is taken and is encrypted using ACM to enhance the security of watermark. First level of LWT is applied over the encrypted watermark. Further, using alpha blending technique HH component of watermark is embedded into the selected block of cover image. To balance the trade-off between robustness and imperceptibility alpha blending technique is used. The steps of proposed color image watermark embedding scheme is discussed in Algorithm 1.

figure a

3.2 Watermark extraction

Let WMRGB represents watermarked image of size m1 × n1. To extract the watermark, firstly cover image and watermarked image is transformed into YCbCr color space. Further, one level of LWT is applied over Y component of cover image and watermarked image. Alpha blending technique is used to extract the watermark from the watermarked image. Finally, Inverse ACM is applied to decrypt the extracted watermark to get original watermark. LWT based watermark extraction steps shown in Fig. 5 is discussed below in Algorithm 2.

figure b

4 Result & discussion

In this section, the result of the proposed scheme is discussed. In this work, seven standard color cover images size 512×512 and a grayscale watermark image of size 64×64 is used to determine the performance of the proposed scheme. Figure 6 depicts the cover image and watermark used in this work. Table 1 depicts the cover image used in the experiment along with its size and extensions. Throughout this study, to balance the trade-off between imperceptibility and robustness of the proposed scheme value of ‘α’ (scaling factor) in (??) and (??) is taken as 0.70.

Fig. 6
figure 6

Cover Images and Watermark

Full size image
Table 1 Features of cover images used in proposed scheme
Full size table

4.1 Imperceptibility analysis

Imperceptibility means the visual consistency of the original image in the presence of a watermark will stay the same as the watermarked image [14, 17]. To evaluate the imperceptibility of the proposed scheme various objective performance metrics like PSNR, MSE, SSIM, SNR, NCC, and MAE of cover and watermarked images are used. Table 2 depicts the imperceptibility of the proposed scheme. The PSNR value of the watermarked image ranges in between 44.1330 to 45.3858. The average PSNR value of the proposed scheme is 44.6506. Whereas the SSIM value ranges in between 0.9959 to 0.9991. SSIM value closes to 1 specifies the high perceptual excellence of watermarked images. NCC values range from 0.9989 to 0.9997. NCC values close to 1 depicts a better result. In Table 2, higher value of PSNR, SNR, SSIM and NCC denotes better imperceptibility. Conversely, lower value of MAE and MSE denotes better imperceptibility.

Table 2 Imperceptibility result of the proposed scheme
Full size table

4.2 Robustness analysis

Robustness is the major metric used to determine the effectiveness of watermarking schemes. The watermarking scheme’s robustness is its ability to withstand the multiple attacks upon it. The watermarked images endured several attacks to test the robustness of our proposed system. In [37] Stirmark benchmark is depicted as a benchmark for watermarking. Here, the attacks are categorized in some classes like scaling, cropping, geometric distortion, etc. Whereas, in [18] these attacks are classified based on their properties. The robustness of the proposed scheme is checked over the image of Lena and grouped in four key groups [31]:

i) noise addition attack ii) image enhancement attack iii) geometric transformation attack iv) compression attack

Further in noise addition attack robustness is have tested over salt & pepper noise, Gaussian noise, speckle noise, and poisson noise. sharpening, histogram equalization, Gaussian filtering, median filtering, and gamma correction are used to test the robustness of enchantment techniques attack. To evaluate the robustness of geometric transformation attacks, cropping, scaling, and rotation attack is used in proposed work. By adjusting the quality factor of the watermarked images, the proposed scheme is tested against JPEG compression. The robustness of the proposed scheme is calculated using metrics like PSNR, NCC, and BER . The attacked images and the recovered watermarks against various attacks are shown in Figs. 7 and 8 respectively, where [a-o] represents, no attack, salt & pepper noise (m = 0,v = 0.05), Gaussian noise (m = 0,v = 0.05), speckle noise, poisson noise, sharpening, histogram equalization, Gaussian filtering, median filtering, Wiener filtering, gamma correction (gamma= 0.5), rotation 10∘, scaling (0.5), cropping 10%, JPEG compression (QF= 90%). Table 3 depicts the PSNR value between the attacked images and watermarked images. From Table 3 its clear that after applying various attacks over watermarked images, it got distorted and the PSNR value decreases. Tables 4, 5, and 6 illustrate the NCC, BER, and PSNR of the original watermark and recovered watermark under various attacks respectively. Experimental results depicted in Tables 4, 5, and 6 shows that that the proposed scheme is robust against attacks. Whereas, in Table 7 NCC, BER and PSNR value of original watermark and the recovered watermark under multiple attacks are depicted. From the result of Table 7 it is evident that the proposed scheme is also robust against combination attacks. The variation of NCC and BER value of watermark and the recovered watermark against salt & peppers noise, Gaussian noise, speckle noise, median filtering, gamma correction attack, rotation attack, and JPEG compression attack with variation in attack parameters is depicted in Figs. 9–16 respectively. From Figs. 9, 10 and 11 it is clear that with the increase in noise density the NCC value decrease and the BER value increases. Whereas, Fig. 12 depicts that with the varying window size from 3×3 to 11×11 there is very small difference in NCC and BER values. Similarly, Fig. 13 depicts the change in NCC and BER value for eight different gamma values. Figure 14 shows the variation in NCC and BER value against varying degrees in both clockwise and anticlockwise. Whereas, the variation in NCC and BER value with the 10 different scaling factors is depicted in Fig. 15. From Fig. 16 and Tables 3, 4 and 5 it is clear that robustness decreased against JPEG compression attack with decrease in QF from (QF= 90) to (QF= 50).

Fig. 7
figure 7

Sample Attacked Lena Images

Full size image
Fig. 8
figure 8

Recovered watermark under different attacks

Full size image
Table 3 PSNR values of attacked image
Full size table
Table 4 NCC values of recovered watermark against various attacks
Full size table
Table 5 BER values of recovered watermark on various attacks
Full size table
Table 6 PSNR values of recovered watermark on various attacks
Full size table
Table 7 NCC, BER and PSNR value on multiple attacks
Full size table
Fig. 9
figure 9

NCC and BER values variation with different salt & pepper noise density

Full size image
Fig. 10
figure 10

NCC and BER values variation with different Gaussian noise density

Full size image
Fig. 11
figure 11

NCC and BER values variation with different speckle noise density

Full size image
Fig. 12
figure 12

NCC and BER values variation for median filtering attack with increase in filter size

Full size image
Fig. 13
figure 13

NCC and BER variation in gamma correction attack with varying gamma value

Full size image
Fig. 14
figure 14

NCC and BER value variation of rotation attack with varying degree

Full size image
Fig. 15
figure 15

NCC and BER of value variation for scaling attack with varying scaling factor

Full size image
Fig. 16
figure 16

NCC and BER value variation for JPEG compression attack with varying QF

Full size image

4.3 Security analysis

Watermark protection is an essential precondition for ensuring the efficient implantation of the watermarking scheme. In this paper, we have used ACM for securing the watermark. To evaluate the security of the watermark performance metrics like NCC, PSNR, SSIM, and BER are used. Table 8 shows the performance compassion of ACM on various iterations. Lower value of NCC, PSNR and SSIM and the higher BER value in Table 8 indicates better security of watermark.

Table 8 Security analysis of watermark
Full size table

4.4 Payload analysis

The payload or embedding capacity is a metric that describes the number of bits of information that can be inserted in the cover image [12, 21, 27]. In this proposed scheme color cover image of size 512 × 512 is used. Also, gray-scale watermark image of size 64 × 64 is used. So, the embedding payload of the proposed scheme is

$$ (64\times64\times8)/(512\times512\times3) = 0.04166\text{ bpp} $$

4.5 Computational complexity

In LWT, the wavelet transform can be computed without allocating the auxiliary memory therefore the LWT based watermarking scheme is memory efficient [24]. The watermark bits are inserted in the chosen blocks, further reducing the expense of the computation.

The average execution time of the proposed LWT based scheme is 0.221461 sec. Table 9 depicts the execution time of the proposed scheme and Kang et al. scheme [12] on the various test images. The average execution time of Kang et al. scheme [12] and Hu et al. scheme [11] is 1.07033 sec and 1.954 sec respectively. Therefore, the proposed scheme is nearly 5 times computationally efficient from Kang et al. scheme [12] and nine-time computationally efficient than Hu et al. scheme [11].

Table 9 Comparison of execution time in (sec) of proposed scheme and Kang et al. scheme [12]
Full size table

4.6 Comparative analysis

In this section comparative analysis of the proposed scheme with the recent state-of-the-art is illustrated. Table 10 demonstrates the imperceptibility comparison between the proposed scheme and some of existing schemes. In Table 11, the robustness comparison with the state-of-the-art of the proposed scheme on the Lena image is performed. From Table 11, it is clear that the proposed scheme is more robust in comparison with the other state-of-the-art schemes. Whereas, Table 12 offers a detailed analysis of the suggested technique with the recent existing schemes.

Table 10 Perceptual Comparison of Lena Image of Size 512 × 512 with state-of-the-art
Full size table
Table 11 NCC comparison of Proposed and state-of-the-art methods Attack
Full size table
Table 12 Performance comparison of the proposed scheme with state-of-the-art
Full size table

5 Conclusion

In this work, a color image watermarking based on the LWT is suggested. As a color image watermarking scheme, the proposed scheme is more appropriate for real-life applications. The suggested approach uses a blending of Alpha to infuse the watermark. Use of LWT makes the proposed technique faster and efficient. Using Arnold transform has improved watermark efficiency. The optimum range of blocks and the alpha blending scheme tackle the trade-off between imperceptibility and robustness. Comprehensive performance analysis of the proposed scheme is conducted including imperceptibility analysis, robustness analysis, security analysis, payload analysis, and computational complexities. The execution time of the proposed scheme is very less as compared to the state-of-the-art. The robustness of the proposed scheme is also tested over various combinations of attacks. Experimental findings suggest the excellence of the proposed scheme is compared to existing schemes. As watermarking is carried out in HH subband, robustness of the proposed scheme against median filtering attack is low. In future, nature based optimization techniques can be used to calculate the optimal α value. Also, robustness against median filtering attack can be improved.