注意力机制和自编码器构造的零水印算法

零水印技术为保护图像版权的有效手段之一. 然而, 现有的许多零水印算法大多采用传统的数学理论进行人工提取特征, 在结合神经网络进行图片特征提取的零水印方向上并没有广泛研究. 目前神经网络在图像特征提取上已经取得了很好的成绩, 充分利用卷积自编码器和注意力机制, 提出了一种用于构造零水印的深度注意自编码器模型(attention mechanism and autoencoder, AMAE). 首先是利用带有注意力的卷积神经网络构建自编码器, 然后对自编码器进行训练; 其次, 利用训练好的编码器输出的特征构造图像的整体特征; 最后, 将获得的特征图进行二值模式处理得到特征二值矩阵, 再与水印图像异或运算得到零水印, 并在知识产权信息数据库进行注册, 零水印一旦注册, 原图像便处于水印技术的保护下. 在训练过程中, 借鉴对抗训练的思想, 对模型进行加噪训练, 这提高了模型的鲁棒性. 实验结果表明, 本文的零水印算法在旋转、噪声和滤波等攻击下, 提取水印图像与原水印图像的归一化系数(normalized correlation, NC)值均超过0.9, 证明了提出算法的有效性和优越性.

Zero-watermarking Algorithm Constructed by Attention Mechanism and Autoencoder

Zero-watermarking technology is an effective means of protecting image copyright. However, most of the existing zero-watermarking algorithms use traditional mathematical theories to extract features manually, and extensive research on zero-watermarking extracting image features with neural networks is still to be conducted. At present, neural networks have achieved favorable results in image feature extraction. A deep attention mechanism and autoencoder (AMAE) model is proposed for constructing zero-watermarks by making full use of a convolutional autoencoder and the attention mechanism. Specifically, an attention-based convolutional neural network is used to construct an autoencoder, which is then trained. Subsequently, the global features of the image are constructed with the features output from the trained encoder. Finally, binary pattern processing of the obtained feature image is conducted to acquire the binary feature matrix. An XOR operation with the image to be watermarked is then performed to obtain a zero-watermark, which is then registered into the intellectual property database. Once the zero-watermark is registered, the original image is under the protection of watermarking technology. During training, the idea of adversarial training is drawn on to train the model with noise, which improves the robustness of the model. The experimental results show that the normalized correlation (NC) values of the extracted watermarked image and the original one to be watermarked both exceed 0.9 under rotation, noise, filtering, and other attacks, which proves the effectiveness and superiority of the proposed algorithm.