A Novel Zero-Watermarking Based Scheme for Copyright Protection of Grayscale Images

Zero-watermarking of digital images is a powerful method with respect to transparency in the watermarked image. However, robustness is still a challenging characteristic for researchers. The proposed method of zero-watermarking provides a novel solution for increasing robustness by obtaining resident features of grayscale image that are robust against common signal processing operations. The proposed solution is based on image scanning to produce NDD (NeighboringDistance Difference) profile. This scheme is used to extract image features for generating redundancy binary profile with the help of image scanning and identification of robust image areas for embedding a binary watermark. Redundant areas from binary profile show perceptually insignificant regions of grayscale image according to human visual system. Resident features from robust areas of image are collected to generate the zero-watermarking binary key image using reversible XOR operation. The binary key is used for extraction of binary watermark. Experimental results of the proposed method have been compared with the results of various zero-watermarking schemes as well as traditional watermarking methods and found much better at slightly higher computational cost. The comparison analysis for testing robustness has been carried out against image processing attacks like Gaussian filtering, block average filtering, motion blur filtering, image resizing, image rotation, image compression and cropping. For each attack maximum correlated watermark from the set of recovered watermarks is selected to evaluate the performance of proposed zero-watermarking scheme. It has been recorded that perfect matching is observed between original and extracted watermarks for a number of signal processing attacks.


INTRODUCTION
D igital image watermarking is being used over two decades for applications such as copyright protection, document identification, authentication, colour or quality hiding, secret data hiding, etc. Conventionally a payload called digital watermark is inserted into the host image. This watermark may be inserted by direct manipulation of image pixels in spatial domain [1][2][3][4] or it may be inserted in transform-domain after applying a discrete transformation on original image [5][6][7][8][9].
Inserting the watermark into host image, faces contradictory requirements of transparency and robustness. On one hand, the inserted watermark needs to be as much transparent as possible to maintain the fidelity of the watermarked image. On the other hand, the inserted watermark needs to be robust against intentional and unintentional attacks on the watermarked image.
While the former requirement constraints to have a low energy watermark, the latter needs the watermark to be strong, consequently having a high energy. The tradeoff required for having robustness and transparency simultaneously has remained a topic of interest for researchers since the inception of the idea of digital image watermarking.
Recently, an alternate scheme by the name of zerowatermarking has become popular [10][11][12][13]. Rather than embedding a watermark in the host image, zerowatermarking applies cryptography by encrypting host image features matrix with digital watermark to produce a watermarking key. This key image with its host image and watermark is then stored in the database of intellectual property for copyright protection. As no watermark is actually embedded into the host image, the watermarked image is the same host image. In other words, the watermarked image is completely transparent.
We present a novel technique based on image scanning and BRF (Binary Robust Feature) profile generation rather than the use of transformations. We exploit the properties of human visual system to use redundant areas for encryption and watermarking key generation.
Experimental results of our scheme are more robust in comparison to the other methods. Rest of the paper comprises of the background and literature review in section II, details of proposed zero-watermarking scheme are given in section III, experimental results and discussions in section IV, finally the paper is concluded in section V. indices in watermarks were embedded by using [14] pixels of the host image in the spatial domain.

BACKGROUND AND LITERATURE REVIEW
Characteristics of histograms and specific color planes of suitable color models are also used for efficient watermarking using spatial domain techniques [15][16][17].
Patchwork [18][19] is also an example of spatial domain watermarking.  [20][21][22]. DWT is another transformation way sort selected level of frequency components in the host image. In [23] a DWT-PCA (Principal Component Analysis) based nonblind color image digital watermarking scheme is proposed.
In [24][25] the HVS (Human Visual System) is exploited using DWT and watermark strength is adjusted with the help of weighting function. DFT based watermarking is another popular scheme because of its robustness against rotational signal processing attacks [9,26].
Zero-watermarking schemes are better than other watermarking schemes in the sense that we can achieve maximum transparency with robustness because no watermark is embedded but encrypted. In [11]  The algorithm is further enhanced in [29] by the same author.  WTG is u max x v max , where u max = (n-2/lm, and v max = (n-2/ ln. This makes the size of WTG same as BRF profile. White squares are paved with watermark as tiles to form a two dimensional WTG g u,v matrix as shown in Fig. 2(b). Right column and bottom row of the mesh tiles is filled with zeroes if smaller than wm x wn.

BRF Profile Generation:
The host Image-I is scanned to find the redundant areas for embedding watermark. Pixels values I i,j are compared with their eight neighbour pixels I ni,j and maximum intensity distance is measured for worst case of redundancy. All image pixels are scanned one by one leaving outer most rows and columns to form the redundancy profile. Here it should be kept in mind that higher intensity distance refers to lower redundancy. This collection forms NDD profile d i,j and may be considered as a greyscale image having size of (m-2) x (n-2) and expressed in Equation (1), where coordinates (i,j) ranges from second to second last row and column of host image respectively.
where I n i,j is n th neighbour of I i,j and D is the function for Euclidian distance between pixels I i,j and I n i,j .
Locations of I i,j and its neighbours I n i,j are shown in Fig. 3(a-c). The NDD profile d i,j is divided into nonoverlapping blocks p u,v of size 3x3. Each block p u,v is used for embedding a single pixel of WTG. In case of binary watermark image, as we are considering here, each bit of WTG will be encrypted with every p u,v block. Maximum value from each p u,v block is taken to identify redundant blocks p u,v . The size of b u,v is u max x v mas which is same as the WTG matrix. Finally, the BRF matrix is generated by comparing block maximum values to a threshold r. If the value of an element is less than r the logical equivalent is marked 1or vice versa. This shows that lmxln block is suitable for embedding if its redundancy value is 1. The value of r is selected on the basis of the JND. The value of JND for human eye is 2.3 in Lab colour space [30]. The equivalent average value on grayscale is used in experiments. Robust profile d i,j and BRF are shown as images in Fig. 3(a-c). Mathematically BRF profile p u,v is expressed in Equation (2). This filtered binary image having the size (512x512) is named the resident image matrix q i,j and it is therefore, robust enough when extracted from the disputed image after tempering by an image processing attack.
Resident image q i,j is divided into non-overlapping blocks q u,v (x,y) with size lmxln leaving outer most rows and columns to make compatible with the blocks of BRF. Since each block will be used for encrypting an individual pixel of watermark therefore a single bit is extracted as resident bit k v,u from each block q u,v (x,y). k u,v is the resident binary matrix RBM and formulated by counting number of ones in blocks q u,v (x,y). If the count of ones is greater than half the number of bits in the block, then resident bit is marked one otherwise it is taken to be zero as calculated in Equation (3).
where(u,v) ranges from (1,1) to (m-2)/lm, (n-2)/ln. Black areas in BRF image are unsuitable for embedding therefore we donot want to use these places in RBM image for robust key generation. Logical AND operation is applied on RBM (k u,v ) and BRF (b u,v ) to filter out undesired locations. RBM and its filtered matrix k u,v are shown in Fig. 4(a-b) and expressed in Equation (4).
Watermarking Key Generation:The filtered RBM k u,v is a binary image. It is robust against most of the intentional or unintentional image processing attacks i.e. it remains unchanged even extracted after intense attacks.
Watermarking key can be generated by applying a reversible operation on the filtered RBM image. Exclusive OR is commonly used for reversible binary operations.
So the k u,v is XORed with WTG to get the watermarking key y u,v as formulated in Equation (5).
At the time of extraction, the watermarking key is again XORed with the extracted filtered RBM image to recover the WTG image.

Identification of Disputed Image Using Registered Binary Key
The disputed image might be attacked by several image processing operations performed intentionally or unintentionally. Disputed image is first scanned in the same way as given in Fig. 1 to produce robust resident features k u,v as illustrated in Fig. 5 by a single block of image scanning. FIG. 4(a). RBMk u,v (170X170) FIG. 4(b). FILTERED RBMk u,v (170X170)

FIG. 5. ZERO-WATERMARK IDENTIFICATION SCHEME
Since the registered key is available at the time of identification therefore robust resident matrix k u,v is operated with registered binary key using XOR to recover WTG array. Watermark tiles from recovered WTG might be in distorted state because of image processing attacks.
Therefore, it is further processed to extract out the watermark tile having maximum correlation with original watermark from WTG array as shown in Fig. 6.

EXPERIMENTAL RESULTS AND DISCUSSIONS
where W i,j and W I,j are original and recovered watermarks.

Common Attacks and Comparison with Previous Work:
SSIM (Structural Similarity) and PSNR approaches are most widely used quantitative measures to assess image fidelity. SSIM is more effective for similarity of two different images while PSNR is an effective way to find the noise contents [31].  Fig. 7(a-p). . 7(a) Proposed identification algorithm is applied on noisy images to recover the watermark using their respective zero-watermarking keys. Maximum correlated recovered watermarks from images given in Fig. 7(a-p) are illustrated in Fig. 8(a-p) respectively.
Recovered watermarks are further analysed by using NC for robustness comparison of the proposed algorithm to previous methods presented in [28][29][31][32][33]. NC results of previous schemes are taken from Rani [29] and compiled in Tables1 Results can be analysed with the help of data tables and graphs shown in Fig. 9(a-b) against two different kinds of signal processing attacks.

Analysis of Robustness against Redundancy Threshold:
Elements of NDD profile d i,j are divided into non-overlapping blocks as expressed in Equation (2)  Tang et. al. [34] have also developed such algorithm and evaluated statistical results. They have also achieved 100% similarities. Results of Tang and proposed scheme are compared in Table 4 against Gaussian noise and JPEG compression.
The proposed scheme embeds an array of 5x5 watermark tiles rather a single watermark. Therefore, the recovered binary image has 25 watermarks with noise contents. Fig. 6 shows the extracted noisy watermarks with tile numbers from JPEG compressed Lenna image with quality factor of 10%.

CONCLUSIONS
The presented watermarking scheme is based on the idea of image scanning to explore the suitable spatial locations Finally, a robust zero-watermarking key is generated. Using this key, the array of watermarks is extracted and maximum matched binary watermark can easily be filtered from the array with the help of highest value of NC. That is why the NC values of extracted watermarks are so ideal in our proposed scheme.