一种基于神经网络的小波域音频水印算法

提出了一种基于小波分解和神经网络的数字音频水印算法，先对音频信号进行小波分解，将数字水印嵌入到音频信号的小波域低频重要系数中，同时通过调整多层前馈神经网络的权重逼近原始音频信号与数字水印之间的关系，然后在接收端用训练好的神经网络提取水印。实验结果表明，嵌入水印的音频信号没有明显的听觉失真；经过噪声干扰，低通滤波，有损压缩，重新采样等信号处理后，相关函数检测具有显著的二值分布特征，且相关系数均达到0．72以上，与其他方法相比，该算法在提取水印时无需原始音频信号，具有运算量低和鲁棒性强等优点。

Audio Watermarking with Neural Networks in the Wavelet Domain

A new watermarking scheme based on wavelet decomposition and neural networks is proposed to offer copyright protection to digital audio signals. A digital watermark is embedded into the significant wavelet coefficients of an audio signal. In the process of watermark being embedded, the weights of the neural networks are regulated till the relationship between the host digital audio and the watermark is approximated. Experimental results show that the correlation functions keep distinct delta function distribution and the correlation coefficients are reached above 0 72 after audio signal manipulations such as noise interference, low pass filtering, lossy compression, and temporal re sampling. Compared with the other algorithms, the proposed one performs watermark detection without participation of the original digital audio and also has the advantages of transparency, low computational complexity and high robustness.