Study On Digital Image Watermarking Algorithm

With the rapid progress of digital technology and multimedia network, digital multimedia products such as digital image, audio, video and so on become more and more easy to be issued, propagated and copied in the network, but this kind of convenience is insecure. So copyright protection for digital multimedia products turns to be a crucial issue. The digital watermarking as a new technique of information security meets these needs. Because the digital watermarking can supply a potential solution to copyright protection and data security, so it has attracted serious concern of scholars and business groups at home and abroad from it was proposed in the early 1990s. At present, great achievements have been made on this research, but the practical application of watermarking still meets some problems to be resolved. And this paper places extra emphasis on the algorithm of digital image watermarking, the main content includes the followings:(1)Review the research background, conceptions, features and fundamental principle of digital watermarking, and introduce some classical algorithms(2)Analyze the statistical character of wavelet coefficients in detail sub-bands, apply the GGD (Generalized Gaussian Distribution) mathematical modeling to this statistical character, and introduce the method of estimation of parameters, design the corresponding detector to detect the watermark. In embedding, Based on the JND (Just Noticeable Difference) in DWT domain, a cost function responds the transparency and robustness is confirmed. According to the cost function, the optimum location can be found for every watermark. During the detection, the LOD (Locally Optimum Detection) detector is designed to detect the watermark. By utilizing the proposed approach, a good compromise between the transparency and robustness can be gotten. The experimental results show that the transparency is improved, but this algorithm has expensive computation and is time-consuming.(3)Combine BP neural network and wavelet significant trees, embed the binary text watermark into the significant coefficients, and the embedded watermark can be extracted without original image. Convert RGB color space to YCbCr color space, and decompose the illumination component with the wavelet. The wavelet coefficients at the same scale of different orientations and different scales of similar orientations have correlativity, and the BP neural network can produce an arbitrary non-linear mapping from input to target output, so this correlativity can be described by the trained BP neural network. In embedding, based on the relationships between actual output and target output of BP neural network, the watermark is embedded into wavelet significant trees, and the embedded watermark can be extracted without original image by the opposite program. The experimental results indicate that the transparency and robustness both can be improved with comparisons to other algorithms, but the weaknesses are obvious: It's difficult to confirm the structure of BP neural network, and training BP neural network with strong ability of generalization is time-consuming.(4)Perform binary conversion to the chaotic random, scan the binary chaotic sequence to chaotic matrix with the inverted zigzag scanning, and encrypted the watermark with the chaotic matrix to improve the security of watermark. Then the new relationship among the chose wavelet coefficients can be built using the chaotic matrix. So the watermark will be embedded into the low frequency wavelet coefficients based on this relationship. During watermark detection, the relationship can be obtained by the same algorithm; then the original watermark can be extracted without original image. The experimental results indicate that the robustness is stronger than other algorithms and the complexity of computation is low, but the transparency is inferior in comparisons.