Robustness Modeling And Algorithm Research For Digital Watermarking

Robustness is one of the most important requirements when digital watermarking is applied. Different from encryption, digital watermarking doesn't strictly require distortion-free transmission. Before people's visual or auditory signals come into being, the host data can tolerate some distortion from signal processing. This requires the embedded watermark can be effectively detected after some signal processing have taken place. The detected ability is known as robustness of digital watermarking. To meet the R&D requirements of robust watermarking, the theory about robustness and algorithm design of watermarking are studied.In order to design the robust watermarking algorithm, the description method and evaluation criteria must be created beforehand. In this regard the main results and contributions of this dissertation are as follows:1. Digital watermarking is viewed as a communication process. The paties participating in the watermarking system are seen as random statistic. A mutual information function between the watermak and the observation is defined as a criterion measuring the robustness of watermarking algorithm. The mutual information metric model can be used as a general framework of watermarking system robustness analysis.2. In the framework of mutual information metric model, robustness of spread spectrum watermarking and quantization watermarking are analyzed with the consideration of transparency and robustness requirements.(1) Spread-spectrum watermarking is divided into simple additive spread-spectrum watermarking and improved spread-spectrum watermarking. The calculation formulas of mutual information function are derived to evaluate the robustness of algorithm. In the experiment, spread spectrum watermarking is implemented in discrete cosine transform (DCT) and the statistic Bit Error Rate (BER) is derived against Gaussian distribution noise and JPEG compression. Experiment results show that evaluation conclusion of mutual information method is in accord with empiric BER. The greater the mutual information function is, the more robust watermarking system is, and the smaller corresponding BER is.(2) The performance analysis of another watermarking paradigm, quantization watermarking is given, which includes Quantization Index Modulation (QIM) watermarking and Distortion Compensated QIM (DC-QIM) watermarking. With the use of mutual information measurement model, their robustnesses are analyzed. Simulation experiments are done to validate the model in Gaussian noise, uniform noise and JPEG compression. Evaluation of the robustness based on the mutual information metric model is matched with the empiric BER. Two evaluation results for reliability of system are consistent.3. When digital watermarking is applied, system performance constraints and convenient implementation scheme are the important consideration. In this dissertation watermark embedding position is selected in DCT domain and realization methods of blind watermarking are studied. Based on the result of mutual information measurement, a method to determine the optimal embedding location is given. In view of robustness, compared with low frequency AC coefficient and high frequency AC coefficient, DC coefficient is a better location. And its PSNR or SNR is acceptable limitation of perception.4. Considering the convenience, the blind watermark detection is a good idea. Taking advantage of Independent Component Analysis's (ICA's) blind source separation characteristics, a blind watermark extraction algorithm is designed in DCT domain. In order to obtain the larger WNR, the visual masking threshold parameter is used to modulate the watermark embedding strength. When the original image and the original watermark are unavailable, a legal user can select the watermarked image and the public reference image as inputs to ICA, and then extract the watermark by ICA algorithm. The public reference image is generated by the key image and original image, but it can be delivered by public transmission channel. So it is not Hidding parameter and does not share the watermarking channel. In this sense, it is a blind watermarking sheme.