深度度量学习视频指纹算法

Video Fingerprinting via Deep Metric Learning

在满足鲁棒性、独特性前提下,为了提高视频指纹系统紧凑性,提出一种端到端的深度度量学习视频指纹算法.网络整体框架由权值共享的三分支网络组成,分支网络采用改进的3D残差网络将多层特征融合并进行压缩,实现视频数据到指纹的端到端映射.网络目标函数由度量和分类双损失函数组成,其中,设计的边界约束三元组角度度量损失函数克服了普通三元组损失函数对特征相关性表达不足的问题;分类损失函数弥补了度量损失对样本特征整体分布不敏感的问题.在公开数据集FCVID上对文中算法、传统方法和深度方法进行了大量实验.结果表明,深度度量学习视频指纹算法在鲁棒性、独特性提高的同时紧凑性显著提高.

In order to improve the compactness,an end-to-end video fingerprinting via deep metric learning is proposed while ensuring its robustness and distinctness.The whole framework is composed of weight-sharing triplet networks.The improved 3D residual network is employed to be the main branch,which fuses multi-layer features together and compresses it.This process maps the raw data to compact fingerprints directly.The new designed boundary-constrained triple angle metric loss and classification loss compose the objective function.The new triple loss overcomes deficient expression to feature correlation.The classification loss function remedies the metric loss which is not sensitive to the overall distribution of sample features.A large number of experiments have been carried out on the FCVID set for the proposed algorithms,traditional methods and deep learning methods.The results show that the algorithm enhances compactness significantly while improving the robustness and distinctness simultaneously.