This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.
Digital watermarking is a process of embedding information, i.e., watermark, into a multimedia element to ensure the validity of the signal. The multimedia element may be an image, audio, video or text. The watermark which has been embedded into the multimedia element can later be extracted for security purposes. In general, digital watermarking can be subdivided into various types such as visible and invisible watermarking, robust and fragile watermarking, blind and non-blind watermarking. Besides that, there are two broad categories of watermarking techniques, namely spatial domain watermarking and frequency domain watermarking. Compared to spatial domain techniques, frequency domain watermarking techniques has been proved to be more effective and at the same time achieve the imperceptibility and robustness requirements of digital watermarking algorithms (Al-Haj, 2007). This project focuses on frequency domain techniques, i.e., DCT-based digital watermarking.
Digital watermarking techniques have been used in various applications, for example copyright protection, fingerprint protection, content authentication, broadcast monitoring, copy control, and biometric protection. What we will focus on this project will be how to apply watermarking techniques for copyright protection.
Instead of using a watermark image, a randomly generated binary bit string is used as the watermark.
The performance of the DCT-based techniques is tested against a gray-scale Lena image of 512 x 512 pixels as the host. The imperceptibility criteria of DCT-based are evaluated using both objective and subjective measurement.
Objective measurements of host image are peak signal-to-noise ratio (PSNR), and mean squared error (MSE). Subjective measurement will be human visual system but in this project, we will only use human inspection to evaluate quality of host image.
The objective of this project is:
To study different types of digital image watermarking techniques specifically DCT-based image watermarking technique.
To analyze the performance of DCT-based techniques.
To improve the robustness and imperceptibility requirement of DCT-based techniques.
2.0 Literature Review
Information hiding is a general term encompassing various sub disciplines (refer Figure 1). The term "hiding" means to make information invincible or keep that particular information undetectable in a host such as images, documents. Two of the most important sub disciplines are steganography and watermarking. (Al-Mualla and Al-Ahmad, 2010)
Figure 1: Classifications of Information Hiding Techniques (Adopted from Petitcolas, 1999)
Cryptography refer to the processing of any information into an encrypted form to prevent it from being read by other people who should not read it or to be able to send it through a channel without being eavesdropped. Although the message is encrypted, hacker or any user who are able to get the message while it is transferring might be able to decrypt it. (Chris, 2002)
Steganography is different from cryptography where steganography hides the fact that a communication is even occurred. Figure 2 shows the process flow to a steganographic system. In steganography, original message which is defined as m is embedded into a harmless message c which act as a cover-object. When the message m is embedded into cover object c, a key k which acts as stego-key will be used. Then the resulting message will be embedded into the cover-object c, which the result can be defined as stego-object s. A cover-object is only used to generate stego-object and will be discarded after stego-object generated. In ideal condition the stego-object should be close to the original cover object so that the presence of embedded information inside a message will be undetected.
Figure 2: Process of a Steganography system
Watermarking is a method of hiding information which is very similar with Steganography. Both methods embed information in a cover message with almost no degradation of the cover-object. The difference of watermarking and steganography is that watermarking adds in the requirement of robustness. An ideal steganographic system is to embed a lot of information but with no visible degradation to the cover object. However, an ideal watermarking system will embed a lot of information that could not be deleted or modified without causing the cover object to be entirely unusable. As a negative effect of the requirement of high robustness, a digital host which has undergone a watermarking process will consume more space compare to those which process by steganographic system.
Figure 3 : Watermark system
Requirement of digital watermarking is as following:
A watermarking system is considered to be high imperceptibility when the watermarked host image looks indistinguishable from the original even on the highest quality equipments. Due to the nature of an image where it is useless to anyone if it has been destroyed or highly distorted by the cover image, a watermarking system which can produce higher imperceptible image is considered a better system.
In the second requirement, a watermark must be highly robust. It should have high resistant against distortion either during normal use, or intentional or malicious attacks such as attempt to disable or remove the watermark. Unintentional attacks include common image transformations such as cropping, contrast and brightness enhancement, etc.
Another less important requirement for ideal watermarking system is capacity. A watermarking system is a system which will perform embedding of a useful amount of information onto an original image. The information embedded can range from a single bit until multiple paragraph of text.
Statistical imperceptibility may be the last possible requirement of an ideal watermarking system. Watermarking algorithm must modify the bits of the cover in such a way that the statistics of the image are not modified in any other way which may indicate that a watermark is presence. This requirement might not be important in steganography, but it do play an important role as one of the requirement in watermarking system.
After we perform watermarking process on an image, there are few methods that we can use to evaluate the particular watermarking techniques. Capacity can easily be evaluate by using the number of bits per cover size. Then statistical imperceptibility is calculated by correlation between the original images and watermarked part.
Besides that, to evaluate a watermarking technique, we can also use Peak Signal-to-Noise Ratio (PSNR). PSNR is the ration between the maximum power of a signal and the power of noise which affects the fidelity of its representation. PSNR can only used as a rough approximation of the quality of the watermark.
Least Significant Bit (LSB) Modification
Least Significant Bit Modification (LSB) is a method where watermark will be embedded into the least significant bit of the cover object (C.Katzenbeisser, 1999). LSB substitution is easy to implement but it also brings a lot of drawbacks. LSB substitution may survive image manipulation such as cropping, but any addition of noise or lossy compression will defeat the watermark. A better attack is to set the LSB bits of each pixel to one. Once the algorithm is detected, the embedded watermark will be easily modified by other party.
Basic LSB substitution can be improved by using a pseudo-random number generator to determine the location of the pixels to be used for embedding based on a key. Securities of the watermark will be improved because the chosen pixels are randomly generated. LSB substitution has been proved to be simple, but this method lacks the basic robustness which is a important requirement for watermarking mechanism.
Frequency Domain Techniques
Discrete-Cosine-Transform (DCT) is the classic and most popular domain for image processing. DCT allows images to be broken up into different frequency bands, making it easier to embed watermarking information into the middle frequency bands of the image.
Figure 4: Definition of DCT Regions
In low frequency region, information inside is very significant and the energy of the region is very high. In contrast, high frequency region has insignificant information and low energy. In low frequency, if information is embedded inside, the original image will get distorted easily. If user embed watermark information into the high frequency region of a host image, when image undergo a compression process, some of the content in the region will be discarded without significant quality degradation (Khayam, 2003). So, the middle frequency band is chosen so that the watermarking process would not affect the most important visual part of host image and get the watermark removed during compression and noise attack.
In accordance to Lin and Chen (2010), watermark information should be embedded into the lower frequency of the host image. They proposed to embed lesser watermark bits into each block of host image so that the influence to the host image can be reduced. A pseudo-random sequence technique of watermark bit stream is used to generate the position of the position to embed watermark. Then, binary bits of watermark will be directly embedded into the least significant bit of DCT coefficient. Duplicate copies of watermark bit will be embedded to increase the robustness of the image.
To extract the watermark, least significant bit of DCT coefficient is replaced by watermark bit stream. So, the proposed method is self-extractable.
The proposed method has been tested against three different host image which have 512 x 512 pixels each and has been proved to be robust to various attack, i.e., JPEG compression and image manipulation, where noise is added into an image.
According to Tang and Aoki (1997), they have proposed a DCT-based method to embed watermark into a host image. In the method proposed, watermark bit string will be embedded into the middle-band of the host image. The process of embedding start by calling an original host image D and watermark image W, the embedding process can be derived as: Dw = D (+) W, where (+) is the process of embedding and Dw is the watermarked images. The requirement of this process is both host images and watermarked image must be in the gray-scale mode. The watermark image size should be half of the original host image, so that when the watermark image is embedded, it would not make the watermarked image become perceptible. Besides that, residual image should be same size with that watermarked image so that watermark embedding can be performed easily.
Watermark embedding process
First, the original host image is broken up into 8 x 8 blocks and a DCT-based algorithm is used to choose coefficient of the original host image in Zig-Zag order. Since the resolution of the watermarked image is half of the original host image, the embedding process is simple.
Then, a Differential Pulse Code Modulation (DPCM) is used to embed the watermark into original host image. The data of the modified watermark will be likely to concentrate on the area [-1,1] of the image (refer figure 5), then after embedding process, the residual pattern is obtained. Then for each marked pixels of the permuted watermark, DCT coefficient is modified according to the residual mask, so that the corresponding residual value is reversed. Lastly, DCT coefficients is inversed to obtain the embedded image.
Figure 5 : The watermark embedding frequency domain
Watermark extraction method
The process of watermark extraction requires both the original image and embedded image. DCT algorithm is used against the original image D and embedded image Dw to get the DCT coefficient. Then the middle-band DCT-coefficients is used to produce the residual patterns. Next, exclusive-or (XOR) operation is used on these two residual patterns to obtain a permuted data. Lastly, the permuted data is decoded to get the watermark W.
Figure 6 : Watermarking process
Figure 7 : Watermark extraction process
This method has been tested and is proved that it is valid in embedding a watermark into an image, and it also has been proved to be effective in decreasing mosquito noise.
Hubali and Kanyakumari (2009) have proposed a watermark generation scheme based on histogram of the image and apply to the original image on the transform DCT domain.
The proposed algorithm is a frequency domain watermarking scheme and operate by modifying the DCT coefficients. The watermark is content based which means that the watermark is generated from the host image itself and no external watermark is used. Binary sequence of watermark is generated with spatial domain information but the embedding process is done in transform domain.
Watermark bit string is generated by calculate the mean of histogram of host image. Hm is denoted as the mean of histogram. Then the gray scale threshold of the image is calculated using Otsu's method. Gt will be denoted as the gray level threshold of the image where the number will be between 0 and 1. Next the mean histogram Hm is downscaled by multiplying it with Gt. the value will be denoted as Th. The original image is then divided into block using formula below (1) where N1 and N2 are the number of rows and columns of image.
Then the mean of the blocks is calculated using following formula (2).
Th and Mb(k;1) will then be compared and generate a string of binary sequence. If Mb(k;1) is larger than Th, computed matrix W will be 0, else if Mb(k;1) is smaller than Th, W will be 1.
The watermark W generated from the original image is embedded into the DCT domain of the image itself. Original host image is divided into blocks of 8 x 8 and DCT transformation is apply to each block using equation below.
The mid frequency coefficients which is generated by DCT process above will be altered using formula below.
Watermark generated will be embedded into the middle frequency coefficient so that these coefficients would not be altered significantly when the image undergo some image manipulation process.
To extract a watermark, original image is required. First, the effect of watermark is undone on the suspicious image. Then the watermark pattern is calculated from the image by repeating the step used to generate watermark pattern.
This method has been tested using matlab image processing toolbox and the output shows that there is no great visual distortion on the image after watermark embedding process. Besides that, it also shows that this method is robust against various image processing i.e., JPEG compression.
In accordance to Taheriniaiand Jamzad, two level DCT based digital watermarking techniques is proved to be highly resistant to compression and additive noise, at the same time preserving high PSNR for watermarked images.
"In spread spectrum communications, a narrowband signal is transmitted over a much larger bandwidth such that, the signal energy present in any single frequency is imperceptible" (A.H. Taherinia, unknown).On the other hand, in spread spectrum watermarking schemes, host image is viewed as a communication channel, and the watermark is viewed as a signal to be transmitted. So, the watermark is spreading many samples of the host signal by adding a low pseudo-random noise sequence to them. The embedded watermark sequence is detected by correlating a specific pseudo-random noise sequence with the watermarked signal.
Blocked based DCT is applied on a host image which is of size 512 x 512. Then, for each 8 x 8 image block, only DC coefficient is selected out of the 64 DCT coefficients to embed the watermark. Selected coefficient will then be mapped into a reduced image which called low-resolution approximation image (LRAI); which is formed from the DC coefficients of all transformed blocks of host image(see Figure 8).
Figure 8 : Low-Resolution Approximation Image
Then according to the value of watermark bit which is going to be embedded in each block, a pseudo random noise sequences added to the high frequencies of DCT transform of each 8 x 8 block of LRAI using formula shown below.
In above formula, FH denotes the high band frequencies, k refer to the gain factor, (x,y) is the location of an 8 x 8 DCT block of LRAI L, and Wi is the pseudo random noise sequence according to the value of i of watermark bit pattern. Two separate pseudo random noise sequences are used to represent the bit values of 0 and 1. Furthermore, by choosing two pseudo random noise sequences to be as un-related as possible, the chance of false detection can be reduced significantly.
Then, each block is inverse-transformed to give us watermarked LRAI. Lastly, DC coefficient of LRAI is replaced with their corresponding watermarked ones, IDCT transform of each 8 x 8 block of LRAI is computed and watermarked image will be constructed.
The same pseudo-random noise generator algorithm is seeded with the same key to detect the embedded watermark. Blocked based DCT is applied on the watermarked image. Then for each 8 x 8 image block, only the DC coefficient is selected out of the 64 DCT coefficients. The coefficients will then be mapped into LRAI. The transformed LRAI of watermarked image is now constructed. Then the correlation between both noise patterns and blocks of transformed LRAI is computed. Pattern with higher correlation will be chosen as extracted watermark bit. Presence of watermark is detected by comparing the average correlation coefficient of detected watermark with a predefined threshold. If the average correlation is greater than the threshold defined, watermark is concluded as presence.
The proposed method has been tested on several tests and has been proved to have highest resistance to JPEG compression compared to other well known techniques. Watermark is able to be extracted even the watermarked image has been compressed with quality factor of 1%. Besides that, this method also shows that it is robust with respect to additive Gaussian noise. Moreover, in this method, watermark can be extracted without even having the embedding parameters. The last advantage of this method is it does not cause any significant changes on watermarked image.