本文设计了一种基于极限学习机算法的离散小波变换域视频水印添加方法, 该方法包括水印嵌入和水印提取两个部分。在水印嵌入环节, 首先使用场景切换检测算法实现非重叠帧提取, 然后对非重叠帧的亮度分量进行5级离散小波变换提取第5级低频子边带系数矩阵, 通过系数矩阵构建训练数据集并通过极限学习机进行回归训练, 回归模型的输出矢量与水印子块对系数矩阵进行修正, 最后通过逆小波变换得到嵌入水印的视频帧序列。在水印提取环节, 对嵌入水印的视频帧序列与原始视频帧序列的亮度分量分别进行5级离散小波变换, 通过提取两个低频子边带系数矩阵的差异部分来得到水印子块, 将所有的子块重组便可得到完整水印。一系列实验显示, 本方法在多个指标下表现良好, 对多种攻击均具有鲁棒性。

A discrete wavelet transform domain video watermarking approach based on extreme learning machine algorithm is designed. The approach includes watermark embedding and watermark extraction. In the watermark embedding process, the scene switching detection algorithm is used to realize non-overlapping frame extraction, and then the fifth-order discrete wavelet transform is applied to the luminance component of the non-overlapping frame to extract the fifth-order low-frequency sub-band coefficient matrix. The training data set is constructed by the coefficient matrix and the regression training is performed by the extreme learning machine. The output vector of the regression model and the watermark sub-block are used to correct the coefficient matrix. Finally, the sequence of video frames embedded in watermark is obtained by inverse discrete wavelet transform. In the watermark extraction process, a 5-level discrete wavelet transform is performed on the luminance component of the watermarked video frame sequence and the luminance component of the original video frame sequence, respectively. The watermark sub-block is obtained by extracting the difference portion of the two low-frequency sub-band coefficient matrices. A complete watermark can be obtained by reorganizing all the sub-blocks. A series of experiments show that the proposed approach is robust and extremely efficient.