www.ietdl.org Published in IET Image Processing Received on 15th September 2010 Revised on 3rd March 2011 doi: 10.1049/iet-ipr.2010.0421

ISSN 1751-9659

Robust image watermarking via geometrically invariant feature points and image normalisation I. Nasir1 F. Khelifi2 J. Jiang3 S. Ipson3 1

Department of Electronic and Computer Engineering, Sebha University, Brak PO. Box 68, Libya School of Computing, Engineering and Information Sciences, Northumbria University, Newcastle Upon Tyne, NE2 1XE UK 3 School of Computing, Informatics and Media, University of Bradford, Bradford BD7 1DP, UK E-mail: [email protected] 2

Abstract: The robustness of watermarks to geometric attacks is viewed as an issue of great importance. Indeed, it constitutes one of the most challenging design requirements for watermarks. This study proposes a robust image watermarking scheme using visually significant feature points and image normalisation. In order to tackle the issue of geometric distortions, the authors adopt a feature extraction method based on end-stopped wavelets to extract significant geometry preserving feature points, which are shown to be robust against various types of common signal processing and geometric attacks. These feature points can be used as synchronisation marks between watermark embedding and detection. The watermark is embedded into nonoverlapping normalised circular images, which are determined by feature points. Rotation invariance is achieved via image normalisation. The watermark embedding process is performed by modifying low-frequency coefficients of discrete cosine transform blocks, which are randomly selected using a secret key. Moreover, the security of the scheme is further guaranteed by an image-dependent key. The proposed scheme is blind as the original image is not required at the watermark detection. Experimental results show that the proposed scheme is robust against geometric attacks as well as common signal processing attacks and outperforms related techniques found in the literature.

1

Introduction

With the rapid growth of the internet and digital media technologies over the last decade, visual data such as images and videos can easily be copied, altered and distributed over the internet without any loss in quality. Therefore the protection of the ownership of multimedia data has become a very challenging issue. In watermarking applications, the robustness of watermarks to attacks is essential to the system [1]. In general, these attacks can be classified into two broad categories: signal processing and geometric attacks. While signal processing attacks attempt to reduce the watermark energy, geometric attacks may induce synchronisation errors between the encoder and the decoder of the watermark. As a result, the decoder is no longer able to detect the watermark. Many digital watermarking schemes have been proposed for copyright protection and several watermarking methods have been developed to overcome the problem caused by geometric attacks. These methods can be roughly classified into template-based, invariant transform domain-based, moment-based, histogram-based and feature extractionbased methods. The template-based watermarking methods are based on embedding a template in addition to the watermark to assist the watermark synchronisation in the detection process. This may be achieved using a structured template embedded in 354 & The Institution of Engineering and Technology 2012

the discrete Fourier transform (DFT) domain [2 – 4] or by embedding the watermark several times at different location [5]. However, there is an accuracy problem associated with log-polar mapping of the DFT since the inverse transformation requires image interpolation; moreover, this technique has limitations in terms of robustness since the template can be easily deleted by eliminating peak values [6]. In invariant-transform domain methods [7 – 10], watermarks are embedded in affine-invariant domains such as the Fourier – Mellin transform or log-polar domain to achieve robustness against affine transforms. However, watermarking methods involving invariant domains are usually vulnerable to cropping and they are difficult to implement because of the log-polar mapping [10]. In moment-based watermarking methods [11 – 13], watermarks are embedded into normalised-based moments robust against affine transforms. In [14], the watermark was embedded in an affine-invariant domain by using Zernike moment. Moments-based methods are highly vulnerable to cropping because of the fact that the moments are extracted from all pixels. Indeed, removal of any part of an image will result in a significant distortion of the moment values. Using the fact that image histograms are independent of the positions of pixels, the authors in [15 – 18] presented a histogram-based watermarking approach. However, these approaches suffer from robustness limitations under histogram enhancement and equalisation attacks. IET Image Process., 2012, Vol. 6, Iss. 4, pp. 354 –363 doi: 10.1049/iet-ipr.2010.0421

www.ietdl.org Another way to reduce or remove the synchronisation issue caused by geometric attacks is to extract feature points, which represent invariant references to geometric transformations. These feature points can be used as reference points for both watermark embedding and detection. These techniques are called second generation watermarking methods [19]. Recently, image feature-based watermarking methods have been widely exploited to overcome the watermark synchronisation issue [20 – 23]. In [20], the Harris detector is used to extract feature points, which are combined with a Delaunay Tessellation to define a number of triangular regions for embedding the watermark. The drawback of this method is that extracted features points from the original and attacked images are not matched. Therefore the sets of triangles generated during watermark embedding and detection are different. Furthermore, this method is not robust to most signal processing attacks except JPEG compression [23]. In [21], a Mexican hat wavelet scale interaction method is used to extract feature points and then the watermark is embedded in normalised disks centred at the extracted feature points. In [22], the authors proposed a method in which the feature points are extracted using the scale-invariant feature transform (SIFT). The watermark is embedded into the circular patches generated using the SIFT. A drawback of this method is that polar mapping during the watermark pattern transformation produces interpolation error. In addition, owing to the lengthy time needed to compute the SIFT descriptor and for compensation of alignment error, the application of this method is restricted. In [23], the authors proposed a method similar to that presented in [20], in which the adaptive Harris corner detector is used to extract feature points and the Delaunaytessellation-based triangle matching method is used to reduce the watermark synchronisation problem and resist geometric distortions. Experiments on this method show its weakness to flip attacks since the algorithm cannot match two triangles when one is flipped. Furthermore, the detection process requires the positions of the feature points from the original image to restore the probe image and reduce the synchronisation errors. Therefore extra cost is need for storage. In [24], the authors use the Harris detector to extract the feature points and embed the watermark in circular regions in the spatial domain. The results for this scheme show that its robustness against signal processing and geometric attacks is limited. For instance, the watermark cannot be detected when the watermarked image is attacked by JPEG compression with a quality factor as low as 60%. The analysis of the existing work described above suggests that second generation watermarking methods [20 – 24] using local image feature points as references can provide solutions to resist geometric attacks. However, the feature point extraction techniques adopted by current feature-based approaches, for instance, the Harris detector or the Mexican hat wavelet detector are sensitive to image modification which makes their robustness to specific attacks limited. Moreover, the methods reported in [11 – 13, 20] are not robust to local geometric attacks such as cropping. This is because the normalisation process is applied into the entire image. Indeed, the removal of any part of an image will result in significant distortion of the moment values. In this paper, a robust feature-based image watermarking scheme is proposed. The main contributions include: 1. Combining the advantages of using image normalisation and geometrically invariant feature points, which are extracted using end-stopped wavelets detector. IET Image Process., 2012, Vol. 6, Iss. 4, pp. 354–363 doi: 10.1049/iet-ipr.2010.0421

2. In the proposed scheme, normalisation is applied into subimages rather than the entire image. This is motivated by the fact that the moments depend on all pixels in the normalised image. Indeed, a removal of any part of the image will result in significant distortion of the moment values. 3. Consideration of the local properties of feature points to detect the watermark even when some features are cropped. 4. Watermark is securely randomised based on an imagedependent key before embedding. The rest of this paper is structured as follows. Section 2 describes the feature extraction method used in the proposed scheme. In Section 3, the image normalisation process developed for pattern recognition is briefly reviewed. Section 4 covers the details of our watermark embedding and detection process. Section 5 presents experimental results. Conclusions are drawn in Section 6.

2

Feature extraction

Geometric attacks can induce synchronisation errors between watermark embedding and detection processes. As a result, the watermark might not be found during detection. In order to reduce synchronisation errors during the watermark detection process, we look for reference points for synchronisation that are perceptually significant and can thus resist various types of common signal processing and geometric attacks. 2.1 Feature detector based on end-stopped wavelets Monga and Evans [25] proposed an iterative feature detector to extract significant geometry preserving feature points. The detector determines the feature points by computing a wavelet transform based on end-stopped wavelets obtained by applying the first-derivative of Gaussian operator to the Morlet wavelet. Monga and Evans [25] evaluated the performance of this detector against three commonly used detectors namely the Harris corner detector, the maximally stable extremaly region detector and the Hessian Affine detector and concluded that the feature detector based on end-stopped wavelets is the most stable and robust. Therefore in the present scheme, this detector has been adopted to extract the feature points. The feature detection process can be described in the following steps: 1. For each orientation angle of the filter, the wavelet transform is computed – by the convolution of the image with the end-stopped filter with scale and orientation parameters similar to [25]. It is worth noting that the endstopped filter is nothing but the Morlet wavelet combined with the first derivative of the Gaussian filter [26]. 2. Once the wavelet filtered versions of the image have been obtained, the candidate feature points are identified by looking for the largest magnitude of the wavelet coefficients in a pre-selected neighbourhood. 3. A threshold Tw is applied to eliminate the points in featureless regions of the image and maintain only actual ones In order to determine the regions for each determined feature point for embedding the watermark, a search is carried out within a circular neighbouring region whose radius is set to be R. If the detector response at the centre of the region achieves local maximum, the feature point is selected. Otherwise, it is discarded. To obtain non-overlapping 355

& The Institution of Engineering and Technology 2012

www.ietdl.org regions, the most stable feature points are first selected. Then, any feature points whose corresponding region overlaps with the selected feature points are excluded.

3

normalised image correspond to a normalisation angle equal to 0.

4

Image normalisation

4.1 As mentioned earlier, synchronisation errors between the embedding and the detection of the watermark may be introduced by geometric attacks such as rotation, shearing and translation and although the watermark is still present in the watermarked image, it can no longer be detected. Image normalisation techniques developed for pattern recognition [27] can be used to overcome this problem as suggested in [11]. In the proposed scheme, an image normalisation technique, which is invariant under rotation attacks, is performed on extracted circular images. The normalisation process is defined as follows. Geometric moments mp,q of a greyscale image are defined as  mp,q =

xp yq f (x, y) dx dy

(1)

G

where G is the region of interest. The central moments are defined as 

mp,q =

(x − x)p (y − y)q f (x, y) dx dy

(2)

G

where x and y are the centroids of the image and are defined as x =

m1,0 m , y = 0,1 m0.0 m0,0

(3)

Central moments invariant under rotational transformations are defined as follows

m′30 = a211 a12 m21 + 3a11 a212 m03

(4)

m′21 = a211 a21 m30 + (a211 a22 + 2a11 a12 a21 )m21 + (2a12 a21 a22 + a222 a21 )m12 + a12 a222 m03

(5)

m′12 = a11 a221 m30 + (a221 a12 + 2a11 a21 a22 )m21 + (2a12 a21 a22 + a222 a21 )m12 + a12 a222 m03

m′30 = a321 m30 + 3a221 a22 m21 + 3a21 a222 m12 + a322 m03

(6) (7)

where a11 , a12 , a21 and a22 are affine transformation coefficients calculated from the eigenvectors and eigenvalues of the covariant matrix of the image, as defined in [27]. To perform normalisation against rotation attacks, two tensors are defined as follows t 1 = m′12 + m′30 ,

t 2 = m′03 + m′21

(8)

Then the normalising angle u can be defined as  1 t u = arctan − 2 t

(9)

The normalised image consists of a rotated version of the original image with angle (u). The idea is to make the new 356 & The Institution of Engineering and Technology 2012

Proposed watermarking scheme Watermarking security

Once the feature points have been selected for embedding, the first step consists of randomising the watermark sequence using an image-dependent key. Let mi be the mean value of the ith circular region. A secret integer number Z is combined with a quantised version of (mi); Q(mi) ¼ ⌈a × mi⌉, where a [ [0.5 1] to create another integer value, which will be the seed of a random generator to randomise the watermark sequence. The seed is nothing but the quantised value of the mean added to Z. In this way, the watermark sequence is not extractable unless the key Z is known. Furthermore, the adversary cannot estimate Z, since it requires the knowledge of the quantised mean values which are variable from a circular image to another. Hence, the observation of different circular images does not make it estimable. The randomisation process is illustrated in Fig. 1b. 4.2

Watermark embedding process

The block diagram shown in Fig. 1b provides an overview of the proposed watermark embedding scheme. The watermark is assumed to be of length Nw in binary form. It is denoted by W ¼ {wi, i ¼ 1, . . . , Nw, wi [ (0, 1)}, which is a keybased PN sequence. The private key is shared with the detector to make the decision whether a given watermark is present or not. The watermark is embedded into lowfrequency coefficients of 8 × 8 discrete cosine transform (DCT) blocks, which are randomly selected using a secret key to further enhance the security of the scheme. The proposed watermark embedding process is described as follows. 1. The feature detector based on end-stopped wavelets is applied in order to determine feature points as described in Section 2. These feature points are used as the reference centres of circular sub-images for watermark embedding and detection. 2. For each determined feature point, search within a circular neighbouring region, whose radius is set to be equal to R and then extract non-overlapping circular images for embedding the watermark. 3. The normalisation process is applied to each extracted circular image. As explained in Section 3. The normalised circular image cannot be transferred directly into the frequency domain. Therefore the zero-padding operation could be performed on the normalised circular image or a sub-image could be extracted from the normalised circular image as illustrated in Fig. 2b. In the present method, subimages are extracted from the normalised circular images because zero-padding operation will introduce error after applying the inverse DCT transform method. 4. The DCT is applied to a selected 8 × 8 blocks of the subimages. 5. To achieve robustness against common signal processing attacks, the low-frequency coefficient of the selected DCT block is used to embed the watermark. In the proposed scheme, the DC coefficients are kept unmodified and the first four AC coefficients in zigzag order are selected to embed the watermark. In order to reduce the visual IET Image Process., 2012, Vol. 6, Iss. 4, pp. 354 –363 doi: 10.1049/iet-ipr.2010.0421

www.ietdl.org

Fig. 1 Proposed watermarking scheme a Randomisation of the watermark sequence with an image dependent key b Watermark embedding

M0 or M1 depending on the value of the watermark bit as shown in Fig. 2b by dashed vertical lines. Mi is the centre of interval Li (i ¼ 0, 1). The watermark embedding algorithm can be described as follows: First, the length of embedding intervals for bit 0 and bit 1 is defined as L0 = L1 =

|AC1 | L

(10)

where L0 and L1 are the length of embedding intervals for bit 0 and bit 1, respectively. L represents the number of embedding intervals and |AC1| is the absolute value of the largest DCT coefficient selected from the first four AC coefficients in a zigzag order. Second, to embed a watermark bit 0 or 1, the absolute value of the second largest DCT coefficient |AC2| is quantised to the nearest M0 to embed ‘0’ or to the nearest M1 to embed ‘1’ as follows

Fig. 2 Extraction of a sub-image from a circular region and quantisation of coefficients a Sub-image extraction process b Quantisation using coefficient AC1 for watermark embedding

degradation on the watermarked image, the number of AC coefficients for embedding a watermark bit in each selected DCT blocks is set to 4. This is because of using more coefficients for embedding a watermark bit will cause more distortions of the watermarked image. The watermark embedding process is carried out by quantising the absolute value of the second largest DCT coefficients in the selected DCT blocks to the nearest values IET Image Process., 2012, Vol. 6, Iss. 4, pp. 354–363 doi: 10.1049/iet-ipr.2010.0421

AC∗2 =



M0 , M1 ,

if w = 0 otherwise

(11)

where AC∗2 is the watermarked coefficient. The absolute values of the DCT coefficients AC3 and AC4 are only quantised to the value of AC∗2 if they are greater than the watermarked coefficient AC∗2 as given by AC∗3 = AC∗4 =

 

AC∗2 , |AC3 |,

if AC3 . AC∗2 otherwise

(12)

AC∗2 , |AC4 |,

if AC4 . AC∗2 otherwise

(13)

The signs of the watermarked coefficients are recovered from the original ones. 357

& The Institution of Engineering and Technology 2012

www.ietdl.org The watermarked sub-images are obtained by applying the inverse DCT transform. Finally, the inverse normalised process is applied to each watermarked circular image. The security requirement of the proposed scheme is fulfilled by a secret key, which consists of the seed used to randomise the watermark. Observe that without knowing the secret key, it would not be possible to obtain the watermark sequence. The proposed watermarking scheme can be summarised as follows: 1. Extract feature points from the original image. 2. Select a number of circular regions as described earlier for embedding the watermark where the centre is determined by the feature points. 3. Normalise each circular region. 4. Each normalised circular region is divided into 8 × 8 blocks where each block is DCT transformed. 5. The first four AC coefficients in a zigzag order are used to embed the watermark. 6. Inverse normalisation is applied to each watermarked region. 7. The watermarked image is reconstructed by replacing the original circular regions with the watermarked ones. 4.3

Watermark extraction process

The proposed watermark extraction process is performed without use of the original image (non-watermarked). Hence, the proposed scheme is able to meet the blindness requirements. In the extraction process, the first four steps are similar to that used in the watermark embedding process. A watermark bit is extracted as given below

Wi∗ =



0, 1,

if |AC∗2 | [ L0 if |AC∗2 | [ L1

(14)

where |AC∗2 | is the absolute values of the second largest DCT coefficients of the first four AC coefficients in the selected DCT blocks of size 8 × 8. The AC coefficients are selected in zigzag order, Wi∗ is the extracted watermark bit and L0 and L1 are the embedding intervals for bits 0 and 1, respectively. The extracted watermark sequence is then compared with the original embedded watermark to decide successful detection. Since the watermarked image may be modified intentionally, for example, by the embedded watermark or unintentionally by attacks, the locations of some extracted feature points may be shifted and not determined correctly. As mentioned earlier, the watermark is embedded into all circular images, which are related to extracted feature points. Therefore ownership is proved if the watermark is detected from at least one circular image. The fact that the watermark is embedded into several circular images, rather than just one, makes it very likely to be detected, even after an image is attacked by signal processing or geometric attacks. In the detection process, an appropriate threshold T is used to make a binary decision whether a given watermark is present or not within the image. This threshold is defined by using the false-alarm probability, which may occur in the watermark detection. Since, the extracted watermark bits are independent random variables with the same ‘success’ probability, Binomial trials can be used to calculate the probability of extracted 358 & The Institution of Engineering and Technology 2012

Fig. 3 False-alarm probability for a sub-image

bits, which match the embedded watermark bits as follows Pk =

  n k P (1 − p)n−k k

(15)

where p is success probability of a bit match between the extracted watermark and the embedded watermark bit sequences, n and k denote the number of watermark bits and the number of matched bits, respectively. Based on the assumption that the success probability is 1/2, the falsealarm error probability for each embedding sub-image is defined as given in (16). Pfalse

alarm

=

n    1 k=T

2

n! k!(n − k)!

 (16)

This is the cumulative probability in the case where k ≥ T, where k represents the number of matched bits between the extracted and the original watermark bit sequences and T represents the threshold. The false-alarm probability against various threshold values is shown in Fig. 3. As can be seen, the match between the extracted and embedded binary watermarks sequences in a sub-image corresponds to a false-alarm probability converging to extremely small values (as small as 15 × 1026) when the threshold T is set to 16.

5

Experimental results

The watermark imperceptibility and robustness are evaluated by using 100 different 8 bit greyscale images of size 512 × 512 including well-known standard images such as Lena, Peppers, Baboon and Lake etc. In all experiments, a pseudorandom sequence of size 16 bits is used as a watermark and the radius of each circular image is fixed at 71 pixels and L ¼ 2. 5.1

Watermark imperceptibility

The degree of imperceptibility of the embedded watermark is important to ensure that there is no noticeable visual degradation because of the embedding process. The peak signal-to-noise ratio (PSNR) is adopted to measure the perceptual distortion of the proposed scheme. The structural similarity (SSIM) index is also adopted for assessing the similarity between the original image and the watermarked one. A detailed description of SSIM can be found in [28]. Fig. 4a shows the SSIM values for 100 watermarked images. Ideally, SSIM would be 1 between two identical IET Image Process., 2012, Vol. 6, Iss. 4, pp. 354 –363 doi: 10.1049/iet-ipr.2010.0421

www.ietdl.org

Fig. 4 Experimental results on the visual similarity and illustration of extracted feature points under JPEG compression attacks a SSIM similarity measurement between original and watermarked images b JPEG compression attacks 50% c JPEG compression attacks 20%

images. As can be seen, the SSIM results show high similarities between the original and watermarked images. In the embedding process of the proposed scheme, the distortion of an image depends on the watermark length, the number of quantisation levels for embedding the watermark, the number of extracted sub-images and the number of AC coefficients for embedding a watermark bit in each 8 × 8 DCT block. The larger the number of AC coefficients used for embedding, the more significant the distortion. Also the more the quantisation levels (L) for embedding watermark bits, the smaller the distortion. In the other words, increasing the number of quantisation levels leads to a small change in the AC coefficients. Hence, there is a tradeoff between robustness and imperceptivity. The PSNR values for 100 watermarked images are between 41.78 and 56.29 db. These values are all greater than 30 db, which is the empirically tested threshold value for the image without any perceivable degradation [23]. Table 1 shows the transparency results of the proposed scheme in comparison with those obtained in [21, 23, 24]. These schemes are chosen to evaluate the proposed scheme because of their similarity to the proposed technique in terms of using feature points and embedding the same watermark length

Table 1

PSNR between watermarked and original images (db)

proposed scheme in [21] scheme in [23] scheme in [24]

Lena

Peppers

Baboon

50.82 49.42 43.33 43.21

50.87 56.60 37.62 44.20

49.46 45.70 44.06 43.23

IET Image Process., 2012, Vol. 6, Iss. 4, pp. 354–363 doi: 10.1049/iet-ipr.2010.0421

(16 bits). The results demonstrate that the proposed scheme offers high PSNR values. This is because the proposed scheme embeds a watermark bit in one AC coefficient and only the other two AC coefficients are altered if they do not satisfy the watermark embedding conditions; whereas in [21, 23], a watermark bit is embedded into the magnitudes of two DFT coefficients while in [24], all pixels inside a circular region are altered to embed one watermark bit. As a result more distortions are introduced in the watermarked images. 5.2

Watermark robustness

In order to evaluate the robustness of the proposed watermarking scheme, various common signal processing and geometric attacks were applied to the watermarked images. These signal processing attacks include JPEG-lossy compression, median filtering, Gaussian filtering and Wiener filtering and geometric attacks include rotation, scaling, shearing, linear geometric transformation, translation, row and column removal and cropping attacks. For the JPEG-lossy compression attacks, quality factor varied from 20% (high compression) to 100%. As examples, results for ‘Elaine’ and ‘Opera’ images under JPEG 50% and JPEG 20% attacks, respectively, are shown in Figs. 4b and c, in which the correctly extracted feature points are marked with white circles, whereas the incorrect feature points are marked with grey circles. As can be seen, overall a majority of the watermarked regions can still be correctly detected even when the watermarked image is compressed by JPEG compression with quality factor of 20%. The robustness of the proposed scheme against JPEG compression attack is achieved by embedding the 359

& The Institution of Engineering and Technology 2012

www.ietdl.org watermark into the low-frequency coefficients of the DCT, which are less affected by JPEG compression attack. The robustness of the proposed scheme against geometric attacks was evaluated by applying common geometric attacks, which included rotation, scaling, shearing, linear geometric transformation, translation, row and column removal and cropping attacks. These types of attacks can be applied to watermarked images in order to make the detector lose synchronicity. For the rotation attacks, the watermarked images were rotated by up to 908 and before applying the detection process. Shearing distortion was applied to watermarked images using a factor up to 5% in x- and y-directions. For example, Fig. 5a shows the result of attack on the ‘Lena’ image rotated by 108. Result of shearing attacks on the ‘Lena’ image is shown in Fig. 5b. As can be seen, overall a majority of the watermarked regions can still be correctly detected. This is because the watermark was embedded into a number of local invariant regions of feature points, which are independent of the position of the pixels. Furthermore, the proposed scheme overcomes the synchronisation problem caused by rotation and shearing attacks by normalising the embedding local regions as explained in Section 3. In a cropping attack, a portion of the watermarked image is removed. This leads to unrecoverable loss of some data and can also introduce a synchronisation problem because of the change in pixels locations. As the watermark was embedded into a number of local regions, which are determined independently of pixel locations, the proposed scheme is able to detect the watermark even when the watermarked image is cropped locally or by cropping 25 or 50% of the

whole image. Under a cropping attack some regions that embedded the watermark may be destroyed, but others may remain unchanged. Figs. 5c – e show some results of cropping attacks applied to watermarked images. The performance of the proposed scheme is evaluated against the Tang’s scheme [21]. Tables 2 and 3 depict the results of various common signal processing and geometric attacks, respectively, in comparison with Tang’s scheme [21] on images ‘Lena’, ‘Peppers’ and ‘Baboon’. As shown in Table 2, the proposed scheme performs better than Tang’s scheme under commonly used signal processing attacks, such as JPEG compression down to a quality factor of 30%, median filtering and combination of median filters with JPEG compression attacks. The watermark can also be extracted correctly by the proposed scheme under a variety of geometric attacks, which Tang’s scheme failed to handle as shown in Table 3. For example, rotation attacks with small rotation angles. It can be seen that the watermark can be correctly extracted by the proposed scheme under rotation, row and column removal, cropping, linear geometric transformation, up to 5% shearing in both horizontal and vertical directions and also translation in both directions. The performance of the proposed scheme was also evaluated against Lei’s scheme [24]. The robustness results are illustrated in terms of bit error rate (BER), which is defined as the ratio between the number of incorrectly extracted watermark bits and the length of the watermark sequence. The experimental results show that, the error bits are 1 bit (BER ¼ 0.06) for different attacks, whereas in the scheme presented in [24], the error bits are 6 bits (BER ≤ 0.35). As shown in Table 4, the

Fig. 5 Extraction of feature points from images under geometric attacks a b c d e

Rotation by 108 Shearing x 5%, y 5% Cropping 25% off Cropping 25% off Cropping 50% off

360 & The Institution of Engineering and Technology 2012

IET Image Process., 2012, Vol. 6, Iss. 4, pp. 354 –363 doi: 10.1049/iet-ipr.2010.0421

www.ietdl.org Table 2

Watermark detection results under signal processing attacks (detection rates) Proposed method

no attack JPEG 80% JPEG 70% JPEG 60% JPEG 50% JPEG 40% JPEG 30% median filtering 2 × 2 median filtering 3 × 3 Gaussian filtering 3 × 3 median filtering 2 × 2 + JPEG 90 median filtering 3 × 3 + JPEG 90 Gaussian filtering 3 × 3 + JPEG 90

Table 3

Scheme in [21]

Lena

Peppers

Baboon

Lena

Peppers

Baboon

6/8 5/8 7/8 4/8 4/8 4/8 4/8 2/8 5/8 5/8 2/8 3/8 3/8

5/6 5/6 3/6 4/6 4/6 4/6 3/6 2/6 3/6 4/6 2/6 2/6 3/6

8/11 7/11 8/11 7/11 8/11 7/11 7/11 5/11 8/11 7/11 5/11 9/11 7/11

7/8 6/8 7/8 6/8 5/8 3/8 2/8 1/8 1/8 5/8 2/8 1/8 5/8

4/4 3/4 3/4 1/4 3/4 1/4 0/4 1/4 1/4 1/4 0/4 1/4 2/4

10/11 9/11 11/11 7/11 7/11 5/11 4/11 6/11 2/11 8/11 6/11 1/11 8/11

Watermark detection results under geometric attacks (detection rates) Proposed method

rotation 1 rotation 2 rotation 5 centred cropping 5% off centred cropping 10% off remove 1 row and 5 columns remove 5 rows and 17 columns shearing x 1%, y 1% shearing x 0%, y 5% shearing x 5%, y 5% linear geometric transform 1.007, 0.01, 0.01, 1.012) linear geometric transform 1.010, 0.013, 0.009, 1.011) linear geometric transform 1.013, 0.008, 0.011, 1.008) translation x and y 10 translation x and y 20 centred cropping 5% +JPEG70 centred cropping 10% + JPEG70

Table 4

Scheme in [21]

Lena

Peppers

Baboon

Lena

Peppers

Baboon

2/8 2/8 2/8 5/8 3/8 4/8 4/8 3/8 2/8 2/8 1/8 2/8 3/8 4/8 4/8 4/8 2/8

4/6 4/6 3/6 3/6 4/6 4/6 3/6 3/6 4/6 2/6 4/6 3/6 4/6 5/6 5/6 2/6 2/6

6/11 7/11 6/11 6/11 6/11 7/11 5/11 6/11 5/11 2/11 6/11 5/11 6/11 5/11 5/11 6/11 6/11

3/8 0/8 0/8 2/8 2/8 3/8 0/8 4/8 2/8 1/8 5/8 4/8 4/8 4/8 4/8 2/8 3/8

2/4 0/4 0/4 2/4 2/4 3/4 1/4 1/4 1/4 0/4 1/4 1/4 0/4 2/4 1/4 2/4 2/4

4/11 0/11 3/11 2/11 2/11 6/11 3/11 5/11 3/11 2/11 4/11 4/11 5/11 7/11 5/11 2/11 2/11

Results under some signal processing and geometric attacks (BER) Proposed method

no attack JPEG 80% JPEG 50% JPEG 30% median filtering rotation 5 rotation 20 rotation 45 rotation 90 translation x, y 2 5 translation x, y 2 10 scaling 0.70 scaling 1.20

Scheme in [24]

Lena BER

Peppers BER

Baboon BER

Lena BER

Peppers BER

Baboon BER

0 0 0 0 0 0.06 0 0.06 0 0 0 0 0

0 0 0 0 0 0 0 0.06 0 0 0 0 0

0 0 0 0 0 0 0.06 0.06 0 0 0 0 0

0 0 0 – 0 0 0 0 0 0 0 0.06 0

0 0 0.18 – 0 0 0 0.06 0 0 0 0 0

0 0.31 0.35 – 0 0 0 0.06 0.31 0.06 0 0.35 0

IET Image Process., 2012, Vol. 6, Iss. 4, pp. 354–363 doi: 10.1049/iet-ipr.2010.0421

361

& The Institution of Engineering and Technology 2012

www.ietdl.org watermark can be correctly extracted by the proposed scheme even under JPEG compression with a quality factor as low as 30%, whereas as reported by those authors, the watermark cannot survive under the scheme in [24] when the watermarked image is compressed with a quality factor of 40% or less. The results of testing 100 images against common signal processing and geometric attacks are shown in Tables 5 and 6, respectively. A success rate is defined as the ratio

Table 5 Success rates of the proposed scheme under common signal processing attacks Attack category no attacks JPEG 100 JPEG 80 JPEG 50 JPEG 30 JPEG 20 median filtering 2 × 2 median filtering 3 × 3 Gaussian filtering 3 × 3 Gaussian filtering 5 × 5 Wiener filtering 3 × 3 Wiener filtering 5 × 5

Detection rate, % 100 100 98 98 92 88 90 95 99 99 97 96

Table 6 Success rates of the proposed scheme under geometric attacks Attack category centred cropping 5% centred cropping 10% centred cropping 20% translation 5 pixels translation 10 pixels translation 20 pixels scaling 0.70 scaling 0.90 scaling 1.20 scaling 1.50 1 row and 5 columns removal 5 rows and 17 columns removal 5 rows and 1 column removal 17 rows and 5 columns removal shearing x 1%, y 1% shearing x 0%, y 1% shearing x 1%, y 0% shearing x 0%, y 5% shearing x 5%, y 0% rotation 18 + cropping rotation 28+ cropping rotation 58+ cropping rotation 108+ cropping rotation 308+ cropping rotation 908 linear geometric transform (1.007, 0.01, 0.01, 1.012) linear geometric transform (1.010, 0.013, 0.009, 1.011) linear geometric transform (1.013, 0.008, 0.011, 1.008)

Detection rate, % 99 99 94 100 100 100 93 98 99 99 94 92 92 90 89 91 91 91 90 99 97 94 96 82 95 91

362

& The Institution of Engineering and Technology 2012

92 92

between the number of watermarked images from which the watermark is extracted correctly and the number of test images. The results demonstrate that the proposed scheme is robust to common signal processing and geometric attacks. Finally, the performance of our proposed scheme compared to related work [21, 24] can be justified by the following points: † In the embedding process, the blocks that contain feature points are not used for embedding the watermark bits. As a result, the influence of the embedding process on the feature points is reduced in comparison to [21, 24]. † As explained in Section 4, the proposed scheme embeds a watermark bit by changing one AC coefficient only which results in less distortion on the embedding circular images and hence more accurate normalisation angle can be used at extraction. † Owing to the use of low-frequency coefficients of the DCT for embedding the watermark, better robustness against JPEG compression attack is achieved compared to [24], which embeds the watermark in the spatial domain. † As the normalisation process is applied into sub-images rather than the entire image, robustness against cropping attacks is achieved.

6

Conclusions

This paper presents a robust image watermarking scheme, which is designed to be robust against both signal processing and geometric attacks. In order to resist geometric attacks, visually significant feature points were extracted using the end-stopped wavelets detector. This feature points are used as reference points to eliminate synchronisation errors between watermark embedding and detection. According to the location of the feature points, circular images were extracted, which were watermarked in the DCT domain. Rotation invariance was achieved using an image normalisation technique. The reference image is not required at detection. It has been demonstrated that under most of the commonly used attacks, the proposed watermarking scheme can recover the embedded watermark from a considerable number of circular images. Experimentally, it has been found that the more robust feature points an image has, the better performance the watermarking system achieves.

7

References

1 Peticolas, F., Anderson, R., Kuhn, M.: ‘Attacks on copyright marking system’. Proc. Second Workshop Information Hiding, Portland, 1998, pp. 218–238 2 Pereira, S., O’Ruanaidh, J.J.K., Deguillaume, F., Csurka, G., Pun, T.: ‘Template based recovery of Fourier-based watermarks using log-polar and log-log maps’. Proc. IEEE Int. Conf. Multimedia Computing Systems, 1999, vol. 1, pp. 870– 874 3 Pereira, S., Pun, T.: ‘Robust template matching for affine resistant image watermarks’, IEEE Trans. Image Proc., 2000, 9, (6), pp. 1123–1129 4 Kang, X., Huang, J., Shi, Y.Q., Lin, Y.: ‘A DWT-DFT composite watermarking scheme robust to both affine transform and JPEG compression’, IEEE Trans. Circuits Syst. Video Technol., 2003, 13, (8), pp. 776– 786 5 Kutter, M.: ‘Watermarking resisting to translation, rotation and scaling’. Proc. SPIE Multimedia System Application, 1998, vol. 3528, pp. 423–431 6 Herrigel, A., Voloshynovskiy, S., Rytsar, Y.B.: ‘Watermark template attack’. Proc. SPIE Security and Watermarking of Multimedia Contents III, 2001, vol. 4314, pp. 394–405 7 O’Ruanaidh, J.J.K., Pun, T.: ‘Rotation, scale, and translation invariant digital image watermarking’. Proc. IEEE Int. Conf. Image Processing, 1997, pp. 536– 539 IET Image Process., 2012, Vol. 6, Iss. 4, pp. 354 –363 doi: 10.1049/iet-ipr.2010.0421

www.ietdl.org 8 O’Ruanaidh, J.J.K., Pun, T.: ‘Rotation, scale, and translation invariant spread spectrum digital image watermarking’, Signal Process., 1998, 66, (3), pp. 303–317 9 Zheng, D., Zhao, J., El Saddik, A.: ‘RST-invariant digital correlation’, IEEE Trans. Circuits Syst. Video Technol., 2003, 13, (8), pp. 753 –765 10 Lin, C.Y., Wu, M., Bloom, J.A., Cox, I.J., Miller, M.L., Lui, Y.M.: ‘Rotation, scale, and translation resilient watermarking for images’, IEEE Trans. Image Process., 2001, 10, (5), pp. 767–782 11 Alghoniemy, M., Tewfik, A.H.: ‘Geometric distortion correction through image normalization’. Proc. IEEE Int. Conf. Multimedia Expo, 2000, vol. 3, pp. 1291–1294 12 Alghoniemy, M., Tewfik, A.H.: ‘Geometric invariant in image watermarking’, IEEE Trans. Image Process., 2004, 13, (2), pp. 145–153 13 Dong, P., Brankov, J.G., Galatsanos, N.P., Yang, Y.Y., Davoine, F.: ‘Affine transformation resistant watermarking based on image normalization’, IEEE Trans. Image Process., 2005, 14, (12), pp. 145–153 14 Kim, H.S., Lee, H.K.: ‘Invariant image watermarking using Zernike moments’, IEEE Trans. Circuit Syst. Video Technol., 2003, 13, (8), pp. 766–775 15 Coltuc, D., Bolon, P.: ‘Watermarking by histogram specification’. Proc. SPIE Security Watermarking of Multimedia Contents II, 1999, vol. 3657, pp. 252– 263 16 Coltuc, D., Bolon, P., Chassery, J.M.: ‘Fragile and robust watermarking by histogram specification’. Proc. SPIE Security Watermarking of Multimedia Contents IV, 2002, vol. 4675, pp. 701 –710 17 Roy, S., Chang, E.C.: ‘Watermarking color histograms’. Proc. Int. Conf. Image Processing, 2004, pp. 2191–2194

IET Image Process., 2012, Vol. 6, Iss. 4, pp. 354–363 doi: 10.1049/iet-ipr.2010.0421

18 Xiang, S., Joong, H., Huang, J.: ‘Invariant image watermarking based on statistical features in low-frequency domain’, IEEE Trans. Circuit and Syst. Video Technol., 2008, 18, (6), pp. 777– 789 19 Kutter, M., Bhattacharjee, S.K., Ebrahimi, T.: ‘Toward second generation watermarking schemes’. Proc. IEEE Int. Conf. Image Processing, 1999, vol. 1, pp. 320–323 20 Bas, P., Chassery, J.M., Macq, B.: ‘Geometrically invariant watermarking using feature points’, IEEE Trans. Image Process., 2002, 11, (9), pp. 1014– 1028 21 Tang, C.W., Hang, H.M.: ‘A feature-based robust digital image watermarking scheme’, IEEE Trans. Signal Process., 2003, 51, (4), pp. 950–959 22 Lee, H.Y., Kim, H., Lee, H.K.: ‘Robust image watermarking using local invariant features’, Opt. Eng., 2006, 45, (3), pp. 037002(1–11) 23 Qi, X., Qi, J.: ‘A robust content-based digital image watermarking scheme’, Signal Process., 2007, 87, pp. 1264–1280 24 Li, L., Guo, B.: ‘Localized image watermarking in spatial domain resistant to geometric attacks’, Int. J. Electron. Commer., 2009, 63, pp. 123–131 25 Monga, V., Evans, B.: ‘Perceptual image hashing via feature points: performance evaluation and tradeoffs’, IEEE Trans. Image Process., 2006, 15, (11), pp. 3453–3466 26 Bhattacharjee, S.K., Vandergheynst, P.: ‘End-stopped wavelets for detecting low-level features’. Proc. SPIE Wavelet Applications in Signal and Image Processing VII, 1999, pp. 732–741 27 Pei, S.C., Lin, C.N.: ‘Image normalization for pattern recognition’, Image Vis. Comput., 1995, 13, (10), pp. 711 –723 28 Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: ‘Image quality assessment: from error visibility to structural similarity’, IEEE Trans. Image Process., 2004, 13, (4), pp. 600– 612

363

& The Institution of Engineering and Technology 2012