Digital Speech Watermark Based On Genetic Algorithms And Neural Network Detection

With the development of computer technology and multimedia processing technology, the piracy of the copyright of digital media is widely spread in the digital media market. How to effectively protect the copyright of digital media becomes a serious issue. Nowadays digital watermark is the most effective way to protect copyright. And one of the most important evaluation standard of digital watermark is whether it is robust to the common processing of digital medias.BP neural network can simulate the nonlinear relationship between its input informations and output informations, and it has advantages of a good self-learning, information distributed storage and massively parallel information processing. Recent years it has been widely applied to the research of audio watermark detection. But there are still some limitations of BP neural network. We use Genetic Algorithms to make up for the BP neural network buy using GA to optimize the initial weights of the BP neural network.The optimal weights then be applied as the initial weights of the BP network for the traning of network. We then use the genetic optimized BP network to detect the digital watermark in the digital speech. The experiment data proves that the genetic optimized BP network is more robust than the traditional BP neural network in watermark detection.The research of this paper mainly include:1) The principles, characteristics and models of digital watermark, Neural Networks and Genetic Algorithms are introduced.2) Proposed a wavelet domain watermark system based on BP neural network detection. The watermark is embedded into the low-frequency coefficients after the audio signal being discrete wavelet transformed. Discussed the determination of the structure and parameters of the BP neural network.Then used the network to learn the relationshiops between digital speech and watermark. The trained network then be used to detect watermark embedded in the speech. The experimental results shows that watermark detected by BP network is rubust to most of the common attacks, but there is still room for improvement. And the traning of BP network is slow.3) Optimized the initial weights of BP network by using genetic algorithms. Used the optimized BP network to detect watermark contained in the audio after the digital audio withstood five kinds of attacks, gave the results of the evaluation, and compared with the experimental results of the traditional BP network. Gave conclusions that watermark extracted by optimized BP neural network is more closer to the original watermark in the watermark blind detection applications, and the traning speed of the network is faster.