Robust, Secure and Semi-Blind Watermarking Technique Using Flexible Scaling Factor in Block-Based Wavelet Algorithm

Abstract

Multimedia security has received much attention recently because of the rapid transmission of elements such as text, images, audio, video, software, animation and games. Security is becoming especially critical for content owners concerned about the illegal usage of their original products. Encryption and watermarking are two methodologies for digital applications. Spatial domain and frequency domain watermarking algorithms give very promising results in embedding binary images into the cover images. This paper proposed a new method of semi-blind watermarking technique. The digital images are divided into 4 × 4 blocks and converted into discrete Wavelet transformations (DWTs). The binary image is embedded into each block using the flexible scaling factor method. Experimental results show that the proposed method has higher peak signal to noise ratio (PSNR) and similarity ratio (SR) values compared to the standard Wavelet transformation and block-based Wavelet algorithms. The results prove that the proposed hybrid algorithm is more effective, robust, secure and resistant than DWT and block-based DWT algorithms.

1. Introduction

Over the past decade, with the rapid increase and advancement of technology and the internet, along with the distribution and transformation of digital data (audio, image, video, etc.), in addition to the vast spread of illegal replication of multimedia data, many studies have been conducted and several watermarking schemes have been proposed to protect the copyright information and ownership of the digital data [1,2]. Digital watermark technology is used to secure and protect the owner’s copyright, infringement and manipulation by verifying the authenticity of the digital data [3]. There are many challenges facing digital watermarking including robustness, variations of watermarks, speedy embedding techniques, image types, color and file format. The receiver of the digital data validates the original data along with watermarking using the watermarking technique. Similarly to watermarking there is another methodology used to hide confidential data in a text or image named Steganography [4].

The technology of digital watermarking has many applications in the labeling of user information, certification, distribution, protection and anti-counterfeit of digital media. In terms of copyright, there are many techniques that can be used as copyright protection by embedding imperceptible copyright information in the multimedia data, such as logo, image, text or the serial number of the author. The application of watermarking in different data types results in different techniques. For example, embedding a digital watermarking to an image is known as image watermarking, to a video stream is known as video watermarking or to a pdf or other text documents is known as text watermarking. This technique requires special readers or detectors to detect and validate the data or provide certification [5]. There are three different types of digital watermarking, namely fragile, semi-fragile and robust watermarking. Fragile watermarking is utilized in data forgery detection by identifying any modification to the data. On the other hand, semi-fragile watermarking is utilized in data authentication and recovery by susceptible malicious attacks. Robust watermarking is generally used in copyright authentication due to its ability to resist most attacks [6]. Methods used for extracting watermarks are classified into two techniques, namely blind and semi-blind watermarking. A technique is called blind watermarking if only the secret key is required for extracting the watermark. The technique is called semi-blind watermarking if both the secret key and the watermarked image are used for extracting the watermark [7,8].

The remainder of the paper is organized as follows: Section 2 presents background reading and related work. To provide a deeper understanding of this area, Section 3 introduces methodologies. Section 4 introduces the proposed methodology. Section 5 presents the experimental results. Finally, the paper concludes with Section 6 focusing on the contribution along with a future recommendation. The next section discusses the related work.

2. Related Work

Many methods and techniques have been developed and applied for watermarking. In this section we are mentioning some of the work developed by researchers, mainly related to our work. Singular value decomposition (SVD) and frequency domain were used in several watermarking schemes [9,10,11]. Makbol et al. [12] adopted the redundant Wavelet transform (RDWT)-SVD to obtain a singular values S which is used to embed the watermark. Ernawan et al. [2] proposed a watermarking scheme using RDWT-SVD to determine the embedding regions with less distortion. The proposed scheme used Arnold transform to scramble the watermark image, which sequentially increased the computational cost to accomplish extra security. The author in [4] proposed the development of a Discrete Wavelet Transform (DWT) by embedding the digital watermark in the primitive image, then withdrawing the watermark from the watermarked images using Xilinx System Generator. The proposed model aims to minimize complexity, resource utilization and robustness. Li et al. [13] proposed an FrFt scheme that includes two phases for multiple image encryption, where the first phase generated keys will be used in the second phase to recover the data and keys; a contrary process, reverse FFT decryption, is used.

In [6], the author proposed an improvement to the fragile watermarking using Local Binary Pattern (LBP). The proposed model uses the least significant bits technique to embed the watermarking in the cover image; it embeds 2–3 bit watermark into the rough block and 1–2 bit watermark into the smooth block. The aim of the proposed model was to use the least amount of watermark embedding to achieve a high level of image quality and protection. Kumar et al. in [14] proposed a semi-blind watermarking scheme for digital images using the Bat Algorithm (BA). The results show good peak signal to noise ratio (PSNR) and low normalized correlation (NC) compared to the same hybrid combination with firefly optimization. The Quick Response (QR) code was used as a watermark in [15], which proposes a blind watermarking scrambling the QR through Arnold transform before implanting the hidden information in the host signal. Different metrics were used to check the imperceptibility and to verify the robustness such as similarity index measure (SSIM), NC and Bit Error Rate.

Lichao et al. [16] proposed a blind watermarking system using the deep learning model; two noise layers are proposed to improve the robustness. The watermark is embedded and extracted using the encoder and decoder of the model. This system is close to the PSNR watermarking method and the proposed model shows improvement in robustness, image rotation, clipping and compression. Ahuja et al. [17] proposed a copyright protection model using a blind video watermarking algorithm, the proposed model uses MPEG-2 style to embed the watermarking information by alternating the value of the DCT coefficients during the encoding phase. Three parameters were tested: robustness, perceptibility and elapsed time for embedding the watermark. The results showed some improvement in watermarking attacks, compression attacks, noise attacks and frame-specific attacks. Yıldıray et al. [18] proposed an improvement to strengthen the LSB steganography technique by using a mask to decrease the number of changes on the image by concealing the personal information on the image, allowing an increment in the amount of data to be hidden and minimizing the distortion on the image. The authors in [19] proposed insertion of the watermark image on the edge position of the host or the original image. The proposed model identifies the edge pixel; then these pixels will be converted into binary and the logo image will be embedded and finally the image will be constructed. The proposed model used the edge detection algorithm which detects any local change in the image intensity for inserting the original image into the reference image. The PSNR and Jaccard functions are used for quality measures for images. In [20], the authors proposed an audio blind watermarking algorithm called zero-watermarking seeking the audio frame suitable for the embedded watermark using scrambling encryption based on BTC, Arnold and chaos encryption. The proposed algorithm combines the zero-passing rate and short-time energy characteristics of the audio frame, the aim of the algorithm was to determine the watermarking embedding strength and the average energy of each segment.

Garg et al. [21] proposed a frequency domain transform model to be used as a blind watermarking technique, in this model the Artificial Bee Colony (ABC) was used for optimization purposes. The aim of the ABC technique was to find the best embedding factor for watermarking, this technique required only the watermarked image extraction, therefore it makes the proposed model a blind watermarking technique. The proposed model combined the Discrete Wavelet Transform (DWT) and the Discrete Cosine Transform (DCT), constructing a hybrid method for embedding and providing a better robustness scheme. Finally, the watermark encrypted image was embedded into the cover image. In [22] the authors proposed a watermarking framework based on Deep Neural Network (DNN), where the network outputs the image with a certain watermark. The proposed model validates the ownership of the image by identifying the network that generated the image. A secret key technique was used to control the watermark extraction to improve the security of the watermarking system; the authors proposed an additional step during the neural network training by adding noise to the generated image to resist attacks and improve robustness. Wang et al. [23] proposed an improved encryption algorithm based on the Low-Density Prity Check Code (LDPC) by expanding the key space aiming to reduce the bit error rate and improve the security of the system. The proposed model shows that even if the encryption of the carrier image is illegally cracked, copyright protection still functions by extracting the watermark.

In [24], the authors proposed double encryption using Arnold transform and Memristive Chaotic Oscillators (MCO) as digital watermarking technique. The proposed model used the Arnold Transform and the coefficients resulted from the MCO based on discrete cosine coefficient modification to embed the watermark based on fractal dimensions extracted using Higuchi’s algorithm followed by the pixel transformation. The extreme learning machine was used as a semi-blind extraction of the watermark. Sharma et al. in [25] proposed a dual watermarking model for copyright protection by using the owner’s fingerprint in addition to the signature as watermarks to validate the ownership and authority of the media. The authors proposed two different watermarks; if one is compromised the other one will be a safeguard to the media. For watermark embedding purposes, the coefficient selection procedure is modified in equal proportions in addition to using the DWT, DCT for the embedding block selection procedure. The proposed model operates only on tone/greyscale images and multitoned images.

Yue et al. [26] proposed a novel route to periodic and chaotic bursting by using the delayed pitchfork bifurcations around stable attractors to achieve different bursting dynamics by computing the corresponding characteristic polynomial. The proposed model uses these analytical results for bursting dynamics in the Lü system. This study aims to tackle the issue of complex bursting pattern, the proposed periodic or chaotic bursting oscillations show that the termination of bifurcation delay may lead the trajectory to switch to the different attractors. The authors of [27] used the bifurcation analytical approach to study the bursting dynamical properties of the van der Pol–Duffing system with the parametric delay feedback. Geometric analysis was presented to conclude that the magnitude of the delay is the determinant of dynamical behaviors of spiking modes and the exponent rate of p, except for the time delay, strongly affects Hopf bifurcation bursting. Some novel bursting patters can be observed by using a parametrical delayed feedback which may lead the trajectory to switch to the more complicated attractors. Time delay in the systems to control or regulate bursting motions may be one of the best approaches.

3. Methodologies

In this section, we are going to give an overview of the main methods used for watermarking. There are three types of main frequency domain watermarking algorithms in the literature; discrete cosine transformation (DCT), discrete Fourier Transformation (DFT) and discrete Wavelet transformation (DWT). There are several reasons to use frequency domain in watermarking. The main reason is that the nature of the human visual system (HVS) is better acquired by the spectral coefficients. Another reason is that high frequencies demonstrate the object edges in the image which are very efficient locations to embed watermarks. Frequency domain watermarking algorithms are powerful against image processing attacks such as filtering, contrast adjustment, histogram equalization, adding noise, blurring, compression and brightness.

3.1. Stenography

Steganography is considered the method of hiding a secret message into a carrier signal such as text, audio, image, or video, mainly to avoid the perception of this secret message from unauthorized parties. The important information, in this case, is not the carrier but the hidden message. There are several methods used to apply steganography in all types of carriers but for the sake of the topic of this paper, we have limited our discussion only to techniques applied in images. In this type of steganography, an image serves as a cover and we can embed into it a secret message. The secret message can be embedded because of image processing operations such as lossy compression, quality alteration format conversion, palette modification, filtering, etc. [28,29]. In all cases, the stego object (hidden information) must be small enough not to raise any suspicion that the cover has been altered. Watermarking similarly to Steganography uses the same principles. The only difference is that the message used by watermarking is to identify the ownership of the image; meanwhile, in Steganography the message has no relationship with the cover image; most often this cover lies far from the meaning of the secret message. Furthermore, in watermarking, the secret message should be imperceptible in the meaning that should be perceptually indistinguishable from the original work for a human observer, a very important aspect of the cover.

The stenographic techniques can be implemented in both domains, spatial and frequency domains. The least significant bit (LSB) method, statistical methods, distortion technique or transform domain techniques are some of the main methods to implement steganography. In the spatial domain, most of the techniques alter one or more bits of each byte of the cover image. The most known technique is LSB (least significant bit). It consists of replacing the last bit of each pixel with the message information according to a specific pattern such as column by column, row by row, or in a more advanced approach a random number generator is used for embedding such as ex. zig zag. The application of watermarking in the spatial domain is low cost, low complexity and needs fewer low system requirements. However, these methods lack robustness [28]. Steganographic and Watermarking applications in the frequency domain can be applied in several transformations such as Discrete Wavelet, Discrete Cosine or Discrete Fourier Transform. In all cases, the transformation coefficients can be altered to embed the watermark in the image. Each of these transformations has its strengths and weaknesses. DWT makes use of the characteristic of the human visual system but has higher complexity compared to DCT; DCT transformation offers robustness against compression and DFT has an advantage over the others in terms of robustness during rotation, scaling or translation [28,30].

3.2. Discrete Wavelet Transformation

Discrete Wavelet transform consists of processing the image in the frequency domain where the decomposition will happen in two subspaces. By using digital filters, the information will be separated into low-frequency sub-band and high-frequency sub-band. With this transformation, the image is decomposed into small Wavelets with varying frequencies and limited duration. The Wavelet transform coefficients contain the positions of the information. The original signal can be completely reconstructed by performing a reverse Wavelet transformation on the same coefficients [31].

Based on [29], If we consider a low pass filter equation H (ω) and a high pass filter as G(ω):

$$H\left(\omega \right)=\sum _k{h}_k{e}^{-jk\omega },\mathrm{G}\left(\omega \right)=\sum _k{g}_k{e}^{-jk\omega }$$

(1)

then a signal x[n] can be decomposed recursively as:

$$c_{j-1,k}=\sum _n{h}_{n-2k}{c}_{j,n}\mathrm{and}dj-1,k=\sum _n{g}_{n-2k}{c}_{j,n}$$

(2)

for j = J + 1, J, …J₀ where c_J+1,k = x|k| Z, J + 1 is the high resolution level index and J₀ is the low resolution level index. The coefficients c and d are called the DWT of signal x[n], where c_J0,k is the lowest resolution part of x[n] and d_j,k are the details of x[n] at various bands of frequencies.

The signal x[n] can be reconstructed from its DWT coefficients recursively:

$$c_{j,n}=\sum _k{h}_{n-2k}{c}_{j-1,k}+\sum _k{g}_{n-2k}{d}_{j-1,k}$$

(3)

The classic DWT is implemented through the combination of two filters (low and high) to create four zones of interests LL, LH, HL and HH as per Figure 1.

The low-frequency part includes the main information of the signal while the high-frequency part includes information related to the edge components. From these sub-bands, the LL is the most sensitive part since it contains most of the energy and serves as the base of the image. The other three sub-bands are of interest for watermarking due to the imperfection of the human visual system. Using several layers of decomposition makes the watermarking more robust and difficult to manipulate [32].

Alpha Blending Technique is a popular technique used in image processing and for watermarking. According to the alpha blending technique, the watermark image is obtained by:

$$WMI=k\times \left(LL1\right)+q\times \left(WM1\right)$$

(4)

where:

• WMI = watermarked image;
• LL1 = low frequency approximation of the original image;
• WM1 = watermark;
• k, q = scaling factors for the original image and watermark, respectively.

The alpha blending formula used for watermark extraction is given by:

$$RW=(WMI-k\times LL1)$$

(5)

where:

• RW = recovered watermark;
• LL1 = low frequency approximation of the original image;
• WMI = watermarked image.

3.3. Discrete Cosine Transformation

Discrete Cosine Transform is also a method used often in watermarking. This technique transforms an image from spatial to the frequency domain by separating the image into sub-bands of different regions of interest based on the visual quality of the image [33].

The general equation of the 1D DCT transform is given as:

$$F\left(u\right)={\left(\frac{2}{N}\right)}^{\frac{1}{2}}\sum _{i=0}^{N-1}{A}_{i}cos\left[\frac{\pi u}{2N}(2i+1)\right]f\left(i\right))$$

(6)

With an inverse transform as:

$$F^{-1}\left(u\right)\mathrm{where},A\left(i\right)=\left\{\begin{array}{c}\\ \frac{1}{\sqrt{2}}\mathrm{for}\xi =0\\ 1\mathrm{otherwise}\end{array}\right.$$

(7)

The general equation for 2D transform is given as:

$$F\left(u\right)={(\frac{2}{N})}^{\frac{1}{2}}{(\frac{2}{M})}^{\frac{1}{2}}\sum _{i=0}^{N-1}\sum _{j=0}^{M-1}{A}_i{A}_{j}cos\left[\frac{\pi u}{2N}(2i+1)\right]cos\left[\frac{\pi \nu }{2M}(2j+1)\right]f(i,j)$$

(8)

Its inverse transformation is:

$$F^{-1}(u,\nu )\mathrm{where},A\left(\xi \right)=\left\{\begin{array}{c}\\ \frac{1}{\sqrt{2}}\mathrm{for}\xi =0\\ 1\mathrm{otherwise}\end{array}\right.$$

(9)

• f(i,j) is the intensity of the pixel in row i and column j;
• N is the width of the image and N is the height;
• F(u,v) is the DCT coefficient in row k1 and column k2 of the DCT matrix.

The DCT is applied in sub-parts of the image with a size of 8 × 8 px. In each of these subparts after the DCT transformation, the energy is shifted to the low frequencies on the top left corner leading the rest of the matrix with many low values which leads to the opportunity to neglect them and perform compression [34]. The 64 basic functions of an 8 × 8 matrix are given in Figure 2.

In traditional techniques, the embedding of the watermarking is applied by modifying the DC values. The main challenge is to find an appropriate embedding factor in terms of robustness and invisibility of the watermarking.

3.4. Discrete Fourier Transformation

Discrete Fourier Transform is a transformation that converts a finite sequence of equally spaced samples of a function into a same-length sequence of equally spaced samples of the discrete-time Fourier Transform. In image processing, the samples are pixel values of an image. The watermarking is mostly applied in the middle-frequency bands since the high frequencies represent the main information causing any change to be visually noticed in the spatial domain and the low frequencies are surpassed by the compression techniques. The application of watermarking in middle frequencies makes it robust and imperceptible at the same time [35,36].

The DFT transform can be obtained as per:

$$F(u,\nu )=\frac{1}{MN}\sum _{x=0}^{M-1}\sum _{y=0}^{N-1}f(x,y)e^{-j2\pi (\frac{ux}{M}+\frac{\nu y}{N})}$$

(10)

where f(x,y) represents the M × N grayscale original image and F(u, v) is a DFT coefficient. e–j2π represents the basis function of Fourier Transform.

The inverse Fourier Transform is given as:

$$f(x,y)=\frac{1}{MN}\sum _{u=0}^{M-1}\sum _{\nu =0}^{N-1}F(u,\nu )e^{-j2\pi (\frac{ux}{M}+\frac{\nu y}{N})}$$

(11)

The output of Fourier Transform has both real and imaginary components distributing the information in both magnitude and phase. Since the magnitude contains less information than the phase, it is recommended that the watermarking be embedded by altering the magnitude.

4. Proposed Methodology

Spatial domain watermarking algorithms are robust, secure and adaptive; however, they are not strong against all types of attacks. In this work, frequency domain watermarking algorithms are applied to obtain resistant results in non-blind, semi-blind and blind watermarking for images. DWT, DCT and DFT algorithms have some advantages and disadvantages regarding extracted watermarks and attacks. In this research work, we enhanced the DWT algorithm in semi-blind watermarking in the embedding of binary images. Figure 3 shows the original cover image, binary watermark and watermarked image using the DWT algorithm [37,38].

Dugat et al. [39] proposed a Wavelet-based watermarking algorithm that is robust and secure. The cover image is converted into the DWT with four bands: low frequencies in the LL band and high frequencies in HL, LH and HH bands. Pseudorandom Number (PRN)-based secret information is embedded into the high frequencies of the first level of decomposition. Experimental results show that DWT-based watermarking is very promising; however, embedding only the first decomposition of the cover image is weak for some of the geometric attacks [40].

The Dugat et al. [39] watermark embedding methodology is demonstrated in Algorithm 1.

Algorithm 1: DWT Embedding Algorithm

INPUT: Cover Image I (size is N × N), watermark W (size is N/2 × N/2) and secret key S If a cover image is a color image, convert it into the greyscale image Transform the cover image into the first level decomposition of the discrete Wavelet transformation Select LL coefficients as a low-level frequency and select HL, LH and HH coefficients as high-level frequencies Embed binary watermark W into the cover image I using the following formula. α is a scaling factor that is obtained using several options WI (Watermarked Image) = W + α × I First level of inversion applied to embedded image coefficients to obtain watermarked image WI OUTPUT: Binary image embedded watermarked image WI

In this method, instead of hiding PSN, a binary image is used as a watermark. When PSN is used as a watermarking, the algorithm detects if there is a watermark or not. If embedding a watermark is an image, stamp, logo, or text, the proposed algorithm extracts hidden multimedia elements. The Dugat algorithm is enhanced using block-based DWT in binary images. The block-based algorithm creates 4 × 4, 8 × 8 or 16 × 16 blocks equal to size blocks from the cover image and embeds a binary image into each block using discrete Wavelet transformation. In block-based watermarking, there is greater chance to find the best embedding pixels to hide secret data, especially against attacks.

Algorithm 2 below demonstrates the block-based DWT methodology [41].

Algorithm 2: Block-based DWT Embedding Algorithm

INPUT: Cover Image I (size is N × N), watermark W (size is N/2 × N/2) and secret key S If a cover image is a color image, convert it into the grey scale image Apply 4 × 4 (or another size) of decomposition to the cover image and binary image. Each block size of the cover image and size of the watermark should be the same Transform each block of the cover image into the first level decomposition of the discrete Wavelet transformation Select LL coefficients as a low-level frequency and select HL, LH and HH coefficients as high-level Frequencies in each band Embed each block of the binary watermark W into each block of the cover image I using the following formula. α is a scaling factor that is obtained using several options WI (Watermarked Image) = W + α × I First level of inversion applied to embedded image coefficients to obtain watermarked image WI OUTPUT: Binary image embedded watermarked image WI

Information hiding using Algorithm 2 gives very promising results, especially in vector images. One of the drawbacks is the high CPU time because of the application of the same algorithm to the N × N blocks. Compared to Algorithm 1, this method is more robust and secure. On the other hand, it has more complexity time. The block-based DWT algorithm is weak against cropping, rotation and image scaling attacks. In this research work, Algorithms 1 and 2 are enhanced in order to find the optimal value of α using the flexible scaling factor idea [42].

In Algorithm 3, the flexible scaling factor is applied to find the optimal α value to reach a robust, secure and resistant watermark methodology.

Algorithm 3: Block-DWT Flexible Scaling Factor Embedding Algorithm

INPUT: Cover Image I (size is N × N), watermark W (size is N/2 × N/2) and secret key S If a cover image is a color image, convert it into the grey scale image Apply 4 × 4 (or another size) of decomposition to the cover image and binary image. Each block size of the cover image and size of the watermark should be the same Transform each block of the cover image into the first level decomposition of the discrete Wavelet Transformation Apply flexible scaling factor algorithms to each block. Every block embeds a watermark into the high frequencies using optimal scaling factor values Embed each block of the binary watermark W into each block of the cover image I using the following formula WI (Watermarked Image) = W + α × I First level of inversion applied to embedded image coefficients to obtain watermarked image WI OUTPUT: Binary image embedded watermarked image WI

The flexible scaling factor in this work is defined as follows [43]:

$$FSF=\sum _{i=1}^3\mid A\left(i\right)\mid \frac{{\sum }_{i=1}^{3}A\left(i\right)}{mxn}$$

(12)

where A(i) is the selected neighbor coefficients in block-based DWT, FSF is the flexible scaling factor and mxn is the total number of pixels in the image block.

There are several metrics that can be used in watermarking. Peak signal to noise ratio (PSNR), Mean Square Error (MSE) and Structural Similarity Index (SSIM) are commonly used metrics to measure the quality of the watermarked images or distortions on the images after some processes or attacks. In this work, the similarity ratio (SR) is calculated for each extracted watermark to evaluate the quality of extraction and level of distortion after attacks.

$$MSE=\frac{1}{AxB}\sum _{a=0}^A\sum _{b=0}^B{\left[n(a,b)-m(a,b)\right]}^2$$

(13)

where A and B demonstrate the number of pixels in the watermarked image and original image.

$$PSNR=20x{log}_{10}\frac{MaximumImageA}{\sqrt{MeanSquareError}}$$

(14)

where A is the watermark or distorted image.

$$SSIM=\left[\frac{2a_1{a}_2}{a_1^2+a_2^2}\right]x\left[\frac{2b_{12}}{b_1^2+b_2^2}\right]$$

(15)

$$SIM={\left[Lumiance(a,b)\right]}^i{x=[Contrast(a,b)}^{j}x={\left[Correlation(a,b)\right]}^z$$

(16)

where a is the mean and b is the variance of images. The similarity ratio is the ratio between a number of pixels with the same value and a total number of pixels in the watermark.

5. Experimental Results

The proposed algorithm is compared with other Wavelet algorithms using Lena, Cameraman, Baboon, Boat, Barbara and Peppers images. Figure 4 demonstrates cover images where the embedding process is applied.

Table 1. Peak signal to noise ratio (PSNR) values after embedding binary image to the cover images.

	DWT (dB)	Block-DWT (dB)	FSF Based Block-DWT (dB)
Lena	41.16	43.25	44.41
Cameraman	43.68	43.61	42.19
Baboon	42.27	42.86	42.89
Boat	39.67	40.82	42.86
Barbara	39.82	39.84	43.19
Peppers	40.72	43.43	44.27

shows PSNR values for a watermarked image using Wavelet, block-based Wavelet and flexible scaling factor-based block Wavelet. Results show that the proposed algorithm gives better image quality values in Lena, Baboon, Boat, Barbara and Peppers images, excepting Cameraman [44,45].

Embedding with FSF-based block DWT is compared with the other most common watermarking algorithms DCT and DFT. The cover image is transformed via discrete cosine transformation and the binary logo is embedded into the medium frequencies after zig-zag order. In the DFT algorithm, the binary watermark is embedded into the high magnitudes and is located in the corner of the image after Fourier Transformation.

Table 2. Peak signal to noise ratio (PSNR) values for DCT, DFT and FSF-based Block-DWT algorithms.

	DCT (dB)	DFT (dB)	FSF Based Block-DWT (dB)
Lena	40.87	41.74	44.41
Cameraman	42.02	42.16	42.19
Baboon	42.29	41.27	442.89
Boat	36.94	40.01	42.86
Barbara	40.15	37.91	43.19
Peppers	39.83	41.57	44.27

demonstrates PSNR values of images after DCT, DFT and the proposed methodology. Results shows that the proposed algorithm higher PSNR values than the DCT and DFT algorithms.

There are several measures of good watermarking algorithms such as robustness, security, data capacity and resistance. Image watermarking can be tested with geometric, temporal and statistical attacks. Temporal attacks are usually used for video sequences and 3D applications. Frame dropping, frame averaging and frame swapping are some of the frequently used attacks in the videos. In this work, we applied several statistical image geometric attacks to compare the proposed watermarking algorithm with the other two Wavelet algorithms [46].

Table 3. Peak signal to noise ratio (PSNR) values after common attacks (A: DWT, B: Block DWT, C: FSF Block DWT).

	Lena	Cameraman	Baboon	Boat	Barbara	Peppers	Method
Rotation	31.60	29.67	38.89	31.53	31.79	38.97	A
	31.91	29.12	37.51	30.74	34.97	37.12	B
	34.76	30.61	38.88	32.54	35.61	37.22	C
Cropping	31.63	29.67	38.89	31.53	31.79	38.97	A
	30.85	28.45	34.25	30.19	30.82	36.52	B
	34.57	29.81	35.19	32.27	30.17	38.72	C
Histogram Equalization	28.25	32.49	29.28	28.11	29.67	31.53	A
	29.49	33.25	30.08	28.53	29.12	32.97	B
	32.16	33.96	34.59	29.94	30.09	35.18	C
Contrast Adjustment	31.79	28.25	32.49	29.28	28.11	29.67	A
	34.29	29.07	30.41	25.83	28.61	30.05	B
	34.35	29.53	34.62	28.84	30.07	31.59	C
JPEG 25	38.89	31.53	31.79	28.25	32.49	29.28	A
	35.67	32.88	32.71	24.86	33.18	30.07	B
	38.27	33.62	32.77	31.21	33.46	32.53	C
Gaussian Noise	28.11	29.36	30.85	31.29	23.63	29.54	A
	30.05	29.16	32.59	32.07	27.52	28.42	B
	31.67	32.04	32.84	32.11	29.67	30.09	C
Resizing	32.56	34.52	33.12	30.96	34.59	37.06	A
	34.18	35.07	31.61	32.58	33.81	36.75	B
	34.77	37.91	36.07	34.73	36.18	39.61	C
Salt and Pepper Noise	31.27	30.82	34.19	32.96	34.52	34.04	A
	33.02	32.01	34.92	34.01	35.91	37.03	B
	35.12	33.43	35.23	36.51	36.72	38.61	C

demonstrates PSNR values after the most commonly used attacks rotation, cropping, histogram equalization, contrast adjustment, JPEG 25, Gaussian noise, resizing and salt and pepper noise. PSNR values after attacks are always less than the PSNR value of the watermarked image. There are three algorithms tested with several images. Method A represents the Wavelet algorithm, B represents block-based Wavelet and C represents the proposed algorithm [47,48].

Experimental results show that the proposed algorithm FSF-based block DWT has better results than the other two Wavelet algorithms in the literature. Time complexity in information hiding systems are based on the algorithm used to embed the watermark into the cover image. In addition to that, the pre-processing stage for the cover image is also another important criterion for finding out time complexity. The most common watermarking algorithm DWT has O(N), DCT has O(NlogN) and DFT has O(N2) time complexity. Block-based DWT and proposed FSF-based Block-DWT algorithms have O(NlogN) time complexities.

Table 4. CPU times in embedding and extraction.

	Embedding (s)	Extraction (s)
DCT	3.17	2.24
DFT	4.91	3.59
DWT	1.83	0.97
2 × 2 Block-DWT	2.01	1.14
4 × 4 Block-DWT	2.53	1.19
8 × 8 Block-DWT	2.67	1.31
FSF based Block-DWT	2.69	1.31

shows CPU times in embedding and extraction for binary image watermarking into the grey scale Lena cover image. The block-DWT algorithms require more CPU time for both embedding and extraction than non-block-DWT watermarking.

Table 5. Similarity Ratio (SR) values after common attacks (A: DWT B: Block DWT C: FSF Block DWT).

	Lena	Cameraman	Baboon	Boat	Barbara	Peppers	Method
Rotation	0.841	0.816	0.893	0.796	0.803	0.827	A
	0.863	0.814	0.905	0.826	0.841	0.897	B
	0.869	0.857	0.917	0.829	0.839	0.904	C
Cropping	0.786	0.794	0.762	0.802	0.756	0.812	A
	0.794	0.831	0.754	0.827	0.794	0.808	B
	0.813	0.834	0.791	0.842	0.816	0.819	C
Histogram Equalization	0.834	0.819	0.847	0.851	0.809	0.819	A
	0.847	0.837	0.855	0.869	0.814	0.810	B
	0.881	0.865	0.873	0.894	0.852	0.844	C
Contrast Adjustment	0.871	0.854	0.883	0.901	0.872	0.821	A
	0.879	0.867	0.870	0.914	0.884	0.825	B
	0.891	0.893	0.879	0.914	0.895	0.853	C
JPEG 25	0.902	0.913	0.894	0.921	0.908	0.917	A
	0.909	0.924	0.876	0.917	0.911	0.921	B
	0.918	0.929	0.901	0.930	0.908	0.927	C
Gaussian Noise	0.875	0.881	0.867	0.855	0.902	0.873	A
	0.886	0.896	0.871	0.849	0.914	0.881	B
	0.881	0.905	0.889	0.873	0.913	0.897	C
Resizing	0.924	0.917	0.931	0.922	0.919	0.925	A
	0.929	0.925	0.929	0.934	0.928	0.925	B
	0.946	0.955	0.953	0.951	0.937	0.942	C
Salt and Pepper Noise	0.857	0.882	0.849	0.893	0.901	0.839	A
	0.846	0.880	0.861	0.899	0.895	0.867	B
	0.869	0.901	0.897	0.907	0.906	0.875	C

shows similarity ratios after attacks for Lena, Cameraman, Baboon, Boat, Barbara and Peppers images. SR represents the quality of the extracted watermark. If the ratio is equal to 1, it means the extracted watermark is exactly the same as an embedded watermark. Experimental results show that SR values are greater than 0.8 after attacks. The FSF-based watermarking algorithm is more successful than the other two Wavelet algorithms in the extraction [49].

Figure 5 shows similarity ratio values after attacks such as rotation, JPEG25, Gaussian noise and salt and pepper.

Results are demonstrated in four bands: low frequencies in the LL band and high frequencies in HL, LH and HH bands. In rotation attacks, the SR value is 0.869 in the HH band which is the highest value for Lena. FSF block watermarking algorithm results show that extraction from high frequencies (HH band) is more robust and secure than extraction from low frequencies (LL, HL and LH bands). One of the reasons for this is that high frequencies usually reside on the edge or corner of the object in the image, which is the most suitable place to hide secret information [50].

6. Conclusions and Future Work

There are two efficient techniques for data hiding in images, cryptology and watermarking. Logo, stamp or pseudo-random number can be embedded into the cover image, video or audio using frequency domain algorithms such as DCT, DFT and DWT. Wavelet transformation gives very good results in embedding high frequencies. In the literature, some of the algorithms resist one group of attacks; other algorithms resist another group of attacks. In this research work, we use block-based Wavelet transformation with flexible scaling factors to increase the resistance of the watermarking algorithm. Experimental results show that the proposed method is more efficient, secure, robust and resistant to common attacks. Results are compared with the DWT and block-based DWT algorithm in the literature. Experimental results show that the proposed algorithm gives higher PSNR values for watermarked images and higher SR values for extracted watermarks after attacks such as rotation, salt and pepper, JPEG 25 and Gaussian noise.