Research On Digital Watermarking Techniques Based On Neural Network And SVD

With the growth of network, the problem for the copyright protection of multimedia becomes more and more important, and it has become a pressing issue. Digital watermarking technology, as a main method of providing copyright protection, has become a very heated research topic in the field of information hiding. This thesis is concerned with the applications of the digital watermarking techniques in digital images. Specially, this thesis focuses on the research and implementations of digital watermarking techniques based on neural network and SVD (Singular Value Decomposition). Firstly, the history of digital watermarking development, the states, basic model, characteristics, classification, the current main algorithm and applications of the digital watermarking techniques are introduced. And several directions of development of digital watermarking at next stage are proposed. Secondly, we introduce the related theory for the neural network and SVD. Using image segmentation, a novel watermarking algorithm based on image segmentation and SVD is proposed. In the algorithm, the image can be divided into different regions. And then do the analysis by using two methods, one is using the same quantization step in a region, another is using the adaptive quantization step based on the statistics of blocks in a region. And when the quantization step in a region is the same one, we discussed it when the embedded component is one region of the image or all. The original image is not required for extracting the watermarking, during which those who don't know the secret key can't rightly retrieve the watermark. A novel blind watermarking algorithm which makes use of the characteristic of the neural network and the SVD is presented. In this method, a binary image watermark is embedded into the original image by adaptive quantification process. First, perform wavelet transform on the original image and divide the wavelet coefficients into several blocks. Second, the singular value decomposition of the blocks is computed. The quantization step is adjusted through the neural network according to the statistics of the blocks, and then the watermark is embedded on the singular value of the blocks. The original image is not required for extracting the watermarking. The algorithms above are implemented. Some analyses compared with Paul et al. [6] are drawn out. Experimental results indicate that the present algorithms are robust against JPEG compressing but extremely sensitive to other malicious manipulation such as filtering and random noising.