Abstract
In this research, a novel block-based, integer wavelet transform (IWT) domain digital image watermarking scheme for optimum information embedding using a Fuzzy Rule Based System (FRBS) is proposed. There are three famous conflicting parameters in digital image watermarking, namely, robustness, imperceptibility and payload are linked with each other and a change in one parameter affects others and vice versa. That’s why it is hard to optimize them, jointly because of the inherent non-linearity and in the literature, any pair of parameters is optimized while third is assumed as fixed. In this proposal, this non-linear problem is solved using FRBS by using the logical relationship among three parameters and it consequently suggests the image from the image-bank that may convey the desired payload (capacity) with maximum imperceptibility and robustness. The proposed FRBS is two-fold. Firstly, selection of candidate image blocks from the given image and secondly selection of the candidate coefficients from the already chosen blocks for embedding the desired payload. Images having coefficients greater than a certain threshold are chosen and the payload is embedded. Consequently, the watermarked images are passed through various attacks and the image with maximum robustness is selected. The effectiveness of the proposed scheme is demonstrated through MATLAB simulations and comparison with state-of-the-art techniques.
Introduction
Image watermarking is one of the oldest yet effective techniques for secure data embedding and transmission. Its importance in digital communication systems has been increasing day by day due to several frauds and attacks on the data being transmitted over public channels. Digital information is directly embedded into the cover data for different purposes like ownership authentication, timestamps, name of legitimate receiver and copyright information (a case of watermarking) or intention to transmit a secret message over the common channel without getting intention of unwanted users (steganography). Moreover, digital watermarking allows the documents to be tracked back for ownership verification. The effective watermarking systems should have the properties of robustness, imperceptibility and more payload. Security and complexity are the issues that may be considered in digital image watermarking. There are generally two domains for watermark embedding, namely the spatial domain and the transform domain. Spatial domain embeds the watermark information directly in the pixels of image [1] while spatial domain embeds the watermark information in the transform domain by taking DCT, DWT or some other wavelet transform etc. [2, 3]. Transform domain watermark embedding scheme is presented by exploiting the properties of human visual system (HVS), which specifies the certain regions in the image for embedding the information [4]. Furthermore, chaos base secure watermarking scheme in transform domain is also proposed which is robust against common image processing attacks [5]. Fuzzy logic based watermarking scheme is also presented with the intent of embedding watermark information which is undetectable to the human visual system (HVS). The scheme targets three of five perceptual properties of HVS [6]. Wavelet transform based medical image watermarking scheme using fuzzy logic is also used by exploiting the properties of HVS. HVS search for the pixels with high texture and luminance and then the watermark is embedded for robustness [7].
In [8], authors proposed a novel technique for payload maximizing in digital image watermarking using a fuzzy rule-based system (FRBS). In this regard, the system takes the desired level of imperceptibility and then infers the maximum possible payload that can be embedded within those image regions where FRBS indicates the room for embedding. FRBS takes decision based on HVS. In [9], authors proposed a novel technique for imperceptibility maximization in digital image watermarking using a fuzzy rule-based system (FRBS). In this regard, the system takes the desired amount of payload and then infers to the maximum possible imperceptibility that can be offered by an image out of many in the image bank and then the payload was embedded within those image regions where FRBS indicates the room for embedding. FRBS takes decision based on HVS.
Originally motivated by [10], authors in [11] proposed a reversible and robust watermarking technique using residue number system (RNS) and Product Codes. Reversibility was ensured by RNS while the product codes guarantee the robustness [12]. Authors in [13] and [14] proposed the concept of cubic product codes (CPCs) for robustness in digital image watermarking and calculated its superiority over simple product codes in the same watermarking application. It was concluded that CPC outperforms compared to simple product codes but at the cost of extra processing and code rate payload.
In digital watermarking, the watermark information could be fragile, semi-fragile, robust or hybrid. Fragile watermarks are very sensitive and are used for tamper detection, while the robust watermarks are used to withstand common image processing operations. The watermark having robustness against friendly attacks while fragile against malicious attacks is called semi-fragile while the one having mixed properties is called hybrid watermark [15] Major purposes of the medical images watermarking are authenticity and integrity-control. Here the authenticity is referred to as the measure through which it is ensured that the image source is valid, and image belongs to the right patient. While integrity control is a capability to validate that image has not been tampered by any mean. Because in some cases a little modification in the medical image may cause a wrong diagnosis etc. [16, 17] In [18], the authors focused on integrity control and authentication of patient electronic medical record (EMR). While in [19], the authors focused on three areas in medical image watermarking. That are authentication, data hiding and their combination. In this paper, a DICOM image was selected for experiment and after dividing it into ROI and RONI parts. The major drawback of the technique was fragility of the watermark. Because the EMP is an important information that must not be preserved through the fragile methods that make it vulnerable against many attacks. In [20], the authors proposed an integrity-verification and authentication of a medical image (ultrasound) by considering it from DICOM image bank. ROI was divided by a rectangular region from the RONI area. Then the hash value of the whole image was calculated by means of hash function SHA256. For sake of enhancing the security two secret keys were used. One as hash key and other as embedded watermark key. Eventually, the watermark was embedded in the least significant bits (LSBs) of RONI area. The received image was assumed authenticated if it possessed a high degree of correlation. Akin to [20], in [21], the authors also selected RONI as watermarking area for authentication purpose. However, the reversibility was not ensured, if the RONI was altered. In [22–24], authors proposed digital watermarking for the medical images in the spatial domain to ensure fragility of the medical image, reversibility and robustness of the embedded watermark. In this regard, the watermark was the patient information that was supposed to be robust while the host medical image was sensitive to any type of friendly and malicious attacks. That seems a genuine requirement in medical images because a minor tampering to the medical image may lead to mis-diagnosis. Whereas, at the same time, saving the watermarking by making it robust is important because it contains the patient information for sake of image authentication that whether the undergoing image belongs to the right patient or not.
Based on the literature review, it can be deduced that regardless of the type of image (natural or medical), the three parameters of the digital image watermarking cannot be jointly optimized, and no effort has made so far, in this regard. Merely two parameters are optimized while third is taken as fixed, constant or ignored. In general, it is observed that transform and spatial domain techniques provide robustness and imperceptibility, respectively but not both.
To address the above cited issue, in this paper, Fuzzy Rule Based System based optimal information embedding in a digital image watermarking perspective is proposed. In this technique the three conflicting parameters, of digital watermarking namely imperceptibility, robustness and payload, shown in Fig. 1, are optimized. An image out of an image bank is selected for secure information communication that offers maximum imperceptibility and robustness for the desired payload.
Digital image watermarking parameters.
The remainder of this paper is organized as follows. System model is introduced in Section 2 with Embedding and extraction phases of proposed scheme. Features and performance indicators of digital watermarking are presented in section 3. Section 4 is dedicated to formulating the fuzzy rule-based system for optimal secure information embedding. In section 5 detailed results of the proposed schemes are presented and section 6 concludes the paper.
Block diagram of the proposed system model for embedding desired payload is given in Fig. 2. Cover image of size NxN is divided into non-overlapping blocks of size 8x8. Inverse Wavelet Transform (IWT) is used to decompose each image block into four sub-bands namely LL (approximate), LH (horizontal), HL (vertical) and HH (diagonal) features where L and H stands for low and high bands respectively. The LL sub-band is approximately located at half the original image while the HH sub band contains the high frequency details of image. On the other hand, the HL and LH convey changes of the image. Since, the payload is to be embedded in middle-band so, the middle-band is selected to calculate the three features of human visual system (HVS) in the image namely, brightness, edge and texture. Mean of each block is computed and the candidate block is indicated using FRBS-1. After selecting the candidate block, FRBS-1 is used again to select the coefficients for embedding C d . The images containing number of coefficients greater than a certain threshold (T h ) are chosen. The selected images are then embedded with the C d and are passed through different attacks. Positions for the candidate block and candidate coefficients are also saved in the location map. The location map is a binary matrix which indicates that which blocks, and which coefficients are used for watermarking. As a part of side information, this location is also sent to the decoder to retrieve the message bits.
Block diagram for watermark embedding.
Figure 3 shows the block diagram for extraction of desired payload (the watermark). The watermarked image is again divided into 8x8 blocks and inverse integer wavelet transform (IIWT) of each block is calculated. By using the location map, candidate blocks are indicated initially and then coefficients are indicated. Embedded payload is then extracted from the LSB’s of indicated coefficients and eventually the correlation may be calculated for further analysis.
Cascaded fuzzy rule based system.
In is generally observed that the uniform areas of the host image are sensitive to the addition of watermark so, only small amount of information can be added in these areas whereas, the edge areas are good for embedding greater watermark information. The Human visual system (HVS) has been considered with several phenomenon that permits to adjust the pixel values to dodge the perception [8, 9]. The FRBS has been used here to classify the pixels that possess more payload with less perceptibility. This is done by making use of three HVS characteristics.
HVS characteristics
There are three HVS characteristics being utilized in the proposed technique by exploiting them the FRBS infers certain results. They are: Brightness In general, the brighter background areas are less sensitive than dark ones. So, the FRBS infers more capacity for brighter sides. Edge Edges are more capable of holding watermark information than others [8]. This information is utilized by the FRBS. Texture Texture is somewhat inversely proportional to the visibility of the watermark [25]. This relationship is manipulated by the designed FRBS.
Performance indicators
There are two key performance indicators in digital image watermarking. Peak Signal to Noise Ratio (PSNR) It represents how much the embedded watermark is imperceptible. Higher values of PSNR describe higher level of invisibility/imperceptibility. Because it is the visible difference of original and watermarked image.
Here f and f′ denote the original and watermarked image respectively of size MxN each. Normalized Correlation (Nc) It indicates the degree of robustness of an image. More correlation means more robustness and vice versa.
This section focuses on the design of fuzzy rule-based system. As already described first FRBS takes the image features as input and finds the payload factor alpha. Components of the both FRBS are discussed in succession. For optimum secure information embedding, a joint fuzzy rule-based system (FRBS) is used which is comprised of a two phase FRBS namely FRBS1 and FRBS2. Here a brief description of both is given in Fig. 4.
Block diagram for watermark extraction.
Firstly, an arbitrary cover image of size MxN is given as input to the feature extraction block. Feature extraction block in return provides three features that are brightness sensitivity texture sensitivity and edge sensitivity. FRBS-1 takes the three features of the whole image as input and gives the payload factor alpha (α). FRBS-2 takes payload factor alpha and desired payload as input and returns imperceptibility factor (IF) of the image. That is IF of the image such that the image contains the desired payload. The details pertaining to the components of proposed model are given below.
As mentioned earlier, first FRBS has three input variables namely brightness sensitivity, texture sensitivity and edge sensitivity duly defined in previous section. The input range of brightness and edge sensitivity is between 0-255 and edge sensitivity could either be 0 or 1. Five membership functions are used to cover the input space of brightness sensitivity, two membership functions are used to represent edge sensitivity (low, high) and five membership function for texture sensitivity. These relationship diagrams are shown in Figs. 5–7 respectively.
First input variable brightness sensitivity.
Second input variable edge sensitivity.
Third input variable texture sensitivity.
There is one output variable named payload factor given in Fig. 8. Five membership functions are used to cover the range which is between 0 and 1. As cardinality of rule base is the product of number of membership functions in each input variables, there are total fifty rules in the rule base. As all three features are somewhat directly proportional to the output, the rules are formulated accordingly. The possible values of variable edge sensitivity are 0 or 1, so twenty-five rules are formulated for each case.
The rule format can be expressed as;
IF (T s = Average AND B s = Dim AND E s = 1) THEN (Alpha = Medium)
The rule surfaces are given in Figs. 9 and 10 with Edge sensitivity E s = 1 and 0 respectively. Both figures narrate that higher the values of brightness sensitivity and texture sensitivity, image payload factor is higher. However, this impact is more when edge sensitivity is 1 and less when edge sensitivity is 0, which conforms to the definitions given in previous sections.
In second fuzzy rule-based system (FRBS 2), there are two input variables namely payload factor (alpha) which is outcome of FRBS-1 and second input variable is desired payload C d and output of FRBS-2 is imperceptibility which is desired peak signal to noise ratio (PSNR). As far as the ranges of input variables are concerned, value of first input variable payload factor is between 0 and 1 and five membership functions are assigned to it. Similarly, value of second input variable desired payload C d is also between 0 and 1and five membership functions are assigned to it. These relationships are shown in Figs. 8 and 11 respectively.
Output variable payload.
Rule surface with E s = 1.
Rule surface with E s = 0.
Input variable payload factor.
There is one output variable named imperceptibility factor which is between 30dB to 70dB while there are nine membership functions. This is shown in Fig. 12. As there are five membership functions in each input variable, there are twenty-five rules in the rule base.
Output variable imperceptibility factor.
This section contains the simulation results to validate the proposed scheme. Different images from medical and natural domain are taken with each of size 256x256 pixels as shown in Fig. 15 along with the number of coefficients available for embedding in each image. These coefficients were found based on FRBS based on the three factors given in previous sections. Block threshold and coefficient threshold were taken as 0.3 both. The sequence of simulation is in this way. First, the payload of each image is calculated based on number of coefficients found in each image. Then based on given payload (desired payload), only those images are chosen that can carry the desired payload. After that, out of those selected images that image is selected whose imperceptibility is the highest among them. In case there may exist more than one image with same imperceptibility, then those images are passed through different types attacks and the image found most robust, (with the highest normalized correlation) is taken as selected image.
Rule surface of FRBS-2.
Relationship between payload and imperceptibility.
Image bank with number of coefficients of images, a) Ultrasound image (6775), b) Boat image (28536), c) Barbra image (27423), d) Side X-ray image (11982), e) Pepper image (19200), f) Front x-ray image (7231), g) Skull x-ray image (5623), h) Baboon image (24870), i) The watermark logo.
For simulation purpose the assumed value of desired payload (C d ) is taken as 3KB (kilobytes) sized image logo given in Fig. 15 (i). After considering this assumption, Fig. 16 shows the candidate images that may carry this desired payload. The payload is calculated by the number of coefficients in the image as given by [8]. Number of coefficients is directly proportional to the payload.
The images that can carry the desired payload.
According to Fig. 16 there are only two images in the bank that can carry our desired payload. First the boat image with the number of coefficients 28536 and Barbra image with the number of coefficients 27423. Intuitively, it is apparent that why Boat image can have more payload because it have a significant white area that represents more brightness and higher the brightness more capacity can be inserted, per the discussion in previous sections. Next highest is Barba image, in that image there a major area is comprised of edges. Hence more edges more capacity. Similar observations can be made with the Baboon image the third highest in terms of number of coefficients. Since each coefficient is embedded with one bit on average so these two images can carry our desired payload so that is why these images are selected. Now let us embed the desired payload in these images. This results in Fig. 17. Figure 17 shows the watermarked images with the desired payload of three kilobytes. Now these watermarked images will be introduced to different attacks subsequently.
PSNR of the watermarked images.
Figure 18 shows the results of watermarked Lean image after passing through JPEG attacks with different quality factors (QF).
JPEG compression of barbra image with different QF, a) QF = 85, b) QF = 70, c) QF = 55 d) QF = 40, e) QF = 25 f) QF = 10.
Figure 19 shows the results of watermarked Barbra image after passing through Salt & Pepper noise attacks with different noise variances.
Salt&pepper noise attack on barbra image with different noise variances, a) 0.001, b) 0.002, c) 0.003.
Figure 20 shows the results of watermarked Barbra image after passing through the Gaussian noise attacks with different noise variances with zero mean.
Gaussian noise attack on barbra image with different noise variances and zero mean, a) 0.01, b) 0.02.
Figure 21 shows the effect of JPEG compression attack on the Boat image with different quality factors.
JPEG compression of boat image with different QF, a) QF = 85, b) QF = 70, c) QF = 55, d) QF = 40, e) QF = 25, f) QF = 10.
Figure 12 shows the impact of Salt & Pepper noise on the watermarked Boat image. Different variances of the noise are applied.
Figure 23 shows the impact of Gaussian noise on the watermarked Boat image. Different variances of the noise are applied that is 0.01 and 0.02 respectively in this regard while means of the noise was taken as zero.
Salt&pepper noise attack on boat image with different noise variances, a) 0.001, b) 0.002, c) 0.003.
Gaussian noise attack on boat image with different noise variances and zero mean, a) 0.01, b) 0.02.
Now if we compare the PSNR value of the watermarked Boat and Barra image after JPEG compression attack with different QF. Then the watermarked boat image is passed through JPEG compression attack with QF of 85, 70, 55, 40, 25 and 10, the PSNR comes to 52.36dB which is exactly equal to the PSNR of watermarked Boat image (Fig. 17). Similarly, when the Barbra image is passed through JPEG compression attack with QF of 85, 70, 55, 40, 25 and 10, the PSNR comes to 51.1dB which is exactly equal to the PSNR of watermarked Barbra image given in Fig. 16.
Table-1 demonstrate the Peak Signal to Noise Ratio and Normalized Correlation N c for the Boat image after passing through different attacks. Against JPEG compression attack, the PSNR comes to exactly same level as given in Fig. 16a which shows the high level of imperceptibility. Also, N c comes to 1 which is also maximum and shows the maximum robustness. Against Gaussian noise, proposed scheme also exhibits good N c for the Boat image as well as the PSNR value.
Analysis of Boat image for different attacks
Table-2 demonstrate the Peak Signal to Noise Ratio and Normalized Correlation N c for the Barbra image after passing through different attacks. Against JPEG compression attack, the PSNR comes to exactly same level as given in Fig. 17b which shows the high level of imperceptibility. Also, N c comes to 1 which is also maximum and shows the maximum robustness. Against Gaussian and Salt & Pepper noise, however, the Barbra image is slightly less in terms of Normalized Correlation N c as well as in terms of the PSNR value.
Analysis of Barbra image for different attacks
Table-3 shows the comparison of Boat image and Barbra image in terms of N c and PSNR values against various attacks. For JPEG compression attack of different quality factors, N c comes to 1 for both images. That shows, either Boat image or Barbra image may be selected for JPEG compression attack. Against Salt & Pepper noise of variance 0.002, N c comes to 0.991 for Boat image and 0.9803 for Barbra mage. It means for Salt & Pepper noise, we will select the Boat image as it gives high robustness. Similarly, for Gaussian noise, N c analysis is similar for both images.
Comparison between Boat and Barbra image for Imperceptiblity and Robustness
According to the comparison, given in Table-3, Boat image is the most suitable image that offers maximum imperceptibility and maximum robustness against several attacks for the given payload of watermarking logo of size 3KB as shown in the Fig. 15 (i).
Intuitively, it is also apparent that the Boat image is more capable in terms of capacity, imperceptibility and robustness because of its nature. It contains sharp edges as well as enough observable bright area (clouds and climate), that helps in carrying more capacity by already described Human Visual System factors definition.
To further visualize the comparison of two candidate images, namely Boat image and Barbra image, the results are plotted in the next figures.
Figures 24 and 25 compares both images against the Salt & Pepper noise with different variances and PSNR and Normalized correlation values are demonstrated respectively.
PSNR comparison for Salt & pepper noise.
From the comparison, it is apparent that Boat image is winner in both PSNR and Nc values compared to the Barbra image. So, this test declares that Boat image is winner.
Similarly, Figs. 26 and 27 shows comparison between the candidate images in terms of PSNR and Nc for Gaussian noise attack, respectively.
Nc comparison for Salt & pepper noise.
Nc comparison for zero mean Gaussian noise.
PSNR comparison for zero mean Gaussian noise.
From the comparison almost, similar results can be observed. The Boat image exhibits higher PSNR and Nc values compared to the Barbra image while introducing the zero mean Gaussian noise attack, shown in the figures below, respectively.
In this scheme, a block-based, integer wavelet transform (IWT) domain digital image watermarking scheme for optimum information embedding using a Fuzzy Rule Based System (FRBS) is proposed. The scheme makes use of Human Visual System (HVS) parameters adequately to find the image segments capable of carrying desired payload (capacity) with maximum imperceptibility and robustness. Per the scheme, firstly the capacity of each image from an image bank is computed. Upon embedding the desired capacity in the images with high imperceptibility, the images were offered different attacks. Post attack analysis of the images in terms of normalized correlation provides the most robust image satisfying all the criteria. In this way, the three conflicting parameters of digital image watermarking are optimized for maximum imperceptibility and robustness for a given payload. In future, more other transformations, embedding/extracting methods and hybrid intelligent techniques may be investigated to take the optimization ahead.