基于双向生成对抗网络的图像感知哈希算法

An Image Perceptual Hash Algorithm Based on Bidirectional Generative Adversarial Network

本文提出一种基于双向生成对抗网络(Bidirectional Generative Adversarial Network, BiGAN)的无监督感知哈希生成算法,通过编码网络、生成网络和判别网络间的双向迭代对抗,生成具有较强图像语义特征表示能力的感知哈希码.本算法通过在编码网络和生成网络间添加跳接层网络结构,将原始图像不同维度的特征信息传递到生成网络,提高生成图像语义学习能力与网络收敛速度;同时,在对抗损失中添加均方误差(Mean Sequare Error, MSE)损失,增强生成图像的视觉质量与细节表示能力.最后,基于网络间的多重迭代对抗训练,输出兼备相同来源图像鲁棒性和不同来源图像区分性的高性能图像感知哈希码.本研究首次采用大型图像数据库进行算法性能评价,实验结果表明,基双向生成对抗网络的感知哈希生成算法与当前其他最新研究方案相比具有更强的版权认证与来源检测能力.

An unsupervised perceptual hash generation algorithm based on a bidirectional generative adversarial net-work(BiGAN)is presented.It generates perceptual hash codes with strong image semantic representation capabilities through bidirectional iterative competition between encoding networks,generation networks,and discrimination networks.Moreover,by adding a skip-connection network structure between the coding network and the generation network,different dimensional features of the original image are transformed from the coding network to the generation network,improving the semantic expression ability of the generated image and convergence speed of the network.At the same time,the mean square error(MSE)loss is added to the network adversarial losses to enhance the visual quality and detail representation ability of the generated image.Finally,a high-performance image perception hash code that possesses the robustness of the same source images and the distinguishability of different source images is obtained via multiple iterative adversarial train-ing networks.A large image database is used for the first time to evaluate the performance of perceptual hash generation schemes in this study.Extensive experimental results show that the proposed algorithm has stronger copyright authentica-tion and source detection capabilities than other state-of-the-art schemes.