ABSTRACT
The copyright protection of multimedia content became a critical issue now a days due to the internet does not use secure links and thus information in transmission is vulnerable to interception. Some solution to be discussed to transmit information in such a way that the existence of the multimedia content is unknown to unauthorized users in order to repel their attention. Some important disciplines of information hiding are cryptography, steganography and watermarking. While cryptography is about protecting the content of the text messages, steganography is about concealing their very existence. Watermarking is about hiding multimedia content in other multimedia data. Watermarking and cryptography are closely related, but cryptography scrambles the image so that it cannot be understood. Similar to steganography, watermarking is about hiding information in other image, but difference is that watermark must be somewhat resilience against attempts to remove it. The information hiding approach can be extended to protect the copyright of multimedia content.
Digital image watermarking technique embeds copyright information into multimedia content. In this research work, two techniques are proposed for information hiding. The first approach is the robust digital image watermarking scheme using back propagation neural network in discrete wavelet transform domain. In this approach, the cover image decomposed into red, green, and blue planes, and the blue plane is divided into 8x8 blocks. Human Visual System is insensitive to variations in blue plane, hence blue plane is selected to embed watermark. The fourth level discrete wavelet transform is performed on each block. The discrete wavelet transform process the image sequentially from low resolution to high resolution, hence provides multi-resolution of an image. The bitmap is selected as watermark and embedded into high and middle frequency coefficients of blue plane of cover image using trained back propagation neural network. The advantage of back propagation neural network is that the errors are back propagated so that the weights of the layers are adjusted continuously until getting the desired output. The aim of this network is achieve the balance between the ability to respond correctly to the input pattern that are used for training and the ability to provide good response to the input that are similar. Back Propagation Neural Network has good nonlinear approximation ability and can establish the relationship between original wavelet coefficients and watermarked wavelet coefficients by adjusting the network weights and bias before and after embedding watermark. The robustness of the watermarked image is tested by some normal attacks such as salt&pepper noise, Gaussian noise, JPEG compression, JPEG 2000 compression, median filtering, average filtering, cropping and rotation. The similarity between the embedded watermark and extracted watermark is tested by measuring the normalized correlation coefficient. The proposed blind watermarking algorithm is robust to salt&pepper noise, JPEG compression, average filtering, median filtering and cropping attacks but weak to Gaussian noise, JPEG 2000compression and rotation attacks.
The second approach is the robust digital image watermarking scheme using quantization and dynamic fuzzy inference system. The Dynamic Fuzzy Expert System also known as Dynamic Fuzzy Inference System (DFIS), , is a widely accepted computing framework based on the popular concepts of fuzzy if-then rules, fuzzy set theory and fuzzy reasoning. In this method, the Mamdani type fuzzy method is exploited to determine a valid approximation of quantization step of each DWT coefficient. A rule base is developed to quantize the wavelet coefficients. The basic concept of fuzzy inference system is that variable values are either linguistic variables or words rather than numbers, their use is closer to human intuition. Computing with either words or linguistic variables exploits the tolerance for imprecision and there by reduces the cost of the defined solution. The watermark is embedded into high and middle frequency subbands of the wavelet transformed coefficients of the blue plane of the cover image, using dynamic fuzzy inference system. The imperceptibility of the watermarked image is tested by measuring peak signal to noise ratio. The robustness of the watermarked image is tested by performing different attacks such as salt&pepper noise, Gaussian noise, JPEG compression, JPEG 2000 compression, median filtering, average filtering, cropping, and rotation. However, this approach is weak to median filtering attack. The similarity between inserted watermark and extracted watermark is tested by measuring normalized correlation coefficient.
Finally, the mean square error, the peak signal to noise ratio and normalized correlation coefficients of two methods are compared and results showed that both BPNN and DFIS algorithms provided better results except for median filtering attack.
CHAPTER1
INTRODUCTION
OVERVIEW OF WATERMARKING
Due to the rapid and massive development of multimedia and the widespread use of the internet, there is a need for efficient, powerful and effective copyright protection techniques. A variety of image watermarking methods have been proposed, where most of them are based on the spatial domain or the transform domain. However, in recent years, several image watermarking techniques based on transform domain are developed to provide better robustness and imperceptibility [1].
Digital Image watermarking schemes are typically classified into three categories. Private watermarking which requires the prior knowledge of the original information and secret keys at the receiver. Semi private or semi blind watermarking where the watermark information and secret keys must be available at the receiver. Public or blind watermarking where the receiver must only know the secret keys [2]. The robustness of private watermarking schemes is high to endure signal processing attacks. However, they are not feasible in real applications, such as digital video display copyright protection where the original information may not be available for watermark detection. On the other hand, semi-blind and blind watermarking schemes are more feasible in that situation [3]. However, they have lower robustness than the private watermarking schemes [4]. In general, the requirements of a watermarking system fall into three categories: robustness, visibility, and capacity. Robustness refers to the fact that the watermark must survive against attacks from potential pirates. Visibility refers to the requirement that the watermark be imperceptible to the eye. Capacity refers to the amount of information that the watermark must carry. Embedding a watermark logo typically amounts to a tradeoff occurring between robustness, visibility and capacity.
Digital image watermarking is a kind of technology that embeds copyright information into multimedia content. Any type of image watermarking scheme mainly includes watermark generation, watermark embedding, watermark detection, and watermark attack [5],[31].Digital image watermarking provides copyright protection to an image by hiding appropriate information into cover image to declare rightful ownership of authenticated users [6]. There are four essential factors which are commonly used to determine the quality of the watermarking scheme. They are robustness, imperceptibility, capacity, and blindness [31]. Robustness is a measure of immunity of watermark against attempts to image modification and manipulation like salt&pepper noise, Gaussian noise, JPEG compression, JPEG 2000 compression, median filtering, average filtering, cropping, and rotation etc. [24],[31]. Imperceptibility is the quality of the cover image that should not be destroyed by the presence of watermark. Capacity is the amount of information that can be embedded in to cover image. Extraction of watermark from the watermarked image without the need of original image is referred to as blind watermarking. The non-blind watermarking technique requires the original image to detect and extract the watermark from the watermarked image. The semi-blind watermarking scheme requires the secrete key and watermark bit sequence for extraction. Another categorization of watermarks based on the embedded data is visible or invisible [7], [31].
According to the domain of watermark insertion, the watermarking techniques fall into two categories: spatial domain methods and transform domain methods. Many techniques have been proposed in the spatial domain such as LSB (Least Significant Bit) insertion method, the patch work method and the texture block coding method [8], [31]. These techniques process the location and luminance of the image pixel directly. The LSB method has a major disadvantage that the least significant bits can be easily destroyed by lossy compression [22]. Transform domain method based on special transformations, and process the coefficients in frequency domain to hide the data. Transform domain methods include fast Fourier Transform, Discrete Cosine Transform, Discrete wavelet transform, Curvelet Transform, Counterlet Transform etc. In these methods the watermark is hidden in the high and middle frequency coefficients of the cover image. The low frequency coefficients are suppressed by filtering as noise, hence watermark is not inserted in low frequency coefficients [8]. The transform domain method is more robust than the spatial domain method against compression, filtering, rotation, cropping and noise attack etc.
Yonghong Chen and Jiancong Chen [9] presented a blind image watermarking scheme that embeds watermark messages at different wavelet blocks is presented base on the training of BPNN in wavelet domain. He Xu, Chang Shujuan [10] presented an adaptive image watermarking algorithm which is based on synthetic human visual system characteristic and associative memory function of neural network. N.Chenthalir Indra and Dr. E. Ramraj [11] proposed a system SBS-SOM a neural network algorithm was trained to generate digital watermark values from the image. Chen Yongqinang, Zhang Yanqing, and Peng Lihua [5] presented a DWT domain image watermarking scheme, where genetic algorithm is used to select the wavelet coefficients to embed watermarking bits into the cover image. Samesh Oueslati, et al. [13] presented an adaptive image watermarking scheme based on Full Counter Propagation Neural Network. Maher EL` ARBI, Chokri BEN AMAR and Henri NICOLAS [14] proposed a novel approach to neural network watermarking for uncompressed video in the wavelet domain. Summrina Kanwal Wajid et al [15] proposed the robust and imperceptible image watermarking using Full Counter Propagation Neural Network with lesser complexity and easy apprehension. Cheng Ri.Pia, et.al [16] proposed a new blind watermark embedding/extracting algorithm using the Radial Basis Function Neural Network. Pao-Ta.Yu et al [17] developed watermarking techniques, integrating both color image processing and cryptography, to achieve content protection and authentication for color images. The efficiency of any watermarking method is mainly based on the performance metrics like Peak Signal to Noise Ratio (PSNR) and Normalized Correlation Coefficient (NCC).
MOTIVATION
Based on the above study, it is inferred that the transform domain is better suited for watermarking. The various methods conclude that the wavelet transform sub band resolution is a better method for watermark insertion.
Watermarking in DWT domain has numerous advantages over other transforms; particularly the Discrete Cosine Transform (DCT).Wavelet transformed image is a multi-resolution description of an image. Hence, an image can be shown at different level of resolution and can be sequentially processed from low to high resolution. DWT is closer to the properties of the human visual system than the DCT, as the selection of embedding is flexible by splitting the signal into individual bands. DWT watermarking techniques follow Human Visual System (HVS) characteristics; it is difficult to detect the watermark existence in the cover image. The high frequency area should be avoided for better robustness while the low frequency area should be avoided for low fidelity. Recent work has focused on developing methods for embedding watermarks in the middle frequency, as it provide a good trade-off between robustness and fidelity.
A neural network represents a highly parallelized dynamic system with a directed graph topology that can receive the output information by means of reaction of its state on the input nodes. Back Propagation Neural Network (BPNN) has good nonlinear approximation ability [31]. It can establish the relationship between original wavelet coefficients and watermarked wavelet coefficients by adjusting the network weights and bias before and after embedding watermark.
The Dynamic Fuzzy Inference System (DFIS), also known as Dynamic Fuzzy Expert System, is a widely accepted computing framework based on the popular concepts of fuzzy set theory, fuzzy if-then rules and fuzzy reasoning. Mamdani type DFIS model is exploited in order to determine a valid approximation of a quantization step of each DWT coefficient [31]. Furthermore, the HVS properties are modeled using bi-orthogonal wavelets to improve watermark robustness and imperceptibility.
The efficiency of digital image watermarking process is evaluated based on the properties of perceptual transparency, robustness, computational cost, capacity of data embedding process, recovery of watermark with or without access to the cover image and the tradeoff between capacity, robustness and imperceptibility.
These features have motivated to develop two new methods for watermarking in transform domain using Back Propagation Neural Network (BPNN) and Dynamic Fuzzy Inference System (DFIS).
ORGANIZATION OF THE WORK
The work is organized into seven chapters as follows:
Chapter 1 presents an overview of digital image watermarking and motivation of the current work.
Chapter 2 provides a survey of the related works on Digital Image watermarking, different watermarking algorithms using Image Transforms, Wavelets, with different Neural Networks and Fuzzy Logic approaches, along with their applications, limitations and objectives of the current research work.
Chapter 3 describes the Discrete Wavelet Transform (DWT), Back Propagation Neural Network (BPNN) and Dynamic Fuzzy Inference System (DFIS) and their implementation to perform watermarking.
Chapter 4 concentrates on providing the implementation details of Back Propagation Neural Network (BPNN) in DWT to embed watermark along with experimental results. The efficiency of the algorithm is tested by performing various attacks such as JPEG compression, Median filtering, Cropping, salt& pepper noise and rotation.
Chapter 5 dwells the implementation of the watermarking algorithm in DWT domain using Dynamic Fuzzy Inference System (DFIS). The efficiency of the algorithm is tested by performing various attacks such as JPEG compression, Median filtering, Cropping, salt& pepper noise and rotation.
Chapter 6 compares the Mean Square Error (MSE), Peak Signal to Noise Ratio (PSNR) and Normalized Correlation Coefficient (NCC) of the BPNN and DFIS methods. The proposed methods are designed and implemented using MATLAB 7.8.
Chapter 7 explains the results and conclusions with limitations and future work in detail.
