ID 原文 译文
21395 结果表明，该方法可以滤除大部分的海浪和地磁场噪声，在频段0.4～0.8 Hz范围内海浪的分布明显减少，较大幅度地改善了时域波形，突出了目标产生的磁异常信号，信噪比可提升近11 dB。 The results show that this method can filter out most of the wave and geomagnetic field noise. The wave distribution in the frequency range of 0.4～0.8 Hz is obviously reduced, the wave form in the time domain is greatly improved, the magnetic anomaly signal of the target is highlighted. Signal to noise ratio can be increased by nearly 11 dB.
21396 该方法计算复杂度低，实时性强且易于实现，可为海洋磁异常检测的噪声抑制提供一种有效手段。 The proposed method has the advantages of low computational complexity, strong real-time performance and easy implementation, which can provide an effective measure for noise suppression of marine magnetic anomaly detection.
21397 针对视频分类中普遍面临的类内离散度和类间相似性较大而制约分类性能的问题，该文提出一种基于深度度量学习的视频分类方法。 To solve the common problem of classification performance restriction caused by big intra-classvariations and inter-class similarities in video classification domain, this paper proposes a deep metric learning based video classification method.
21398 该方法设计了一种深度网络，网络包含特征学习、基于深度度量学习的相似性度量，以及分类3个部分。 The proposed method designs a deep network which contains three parts: feature learning, deep metric learning based similarity measure as well as classification.
21399 其中相似性度量的工作原理为：首先，计算特征间的欧式距离作为样本之间的语义距离；其次，设计一个间隔分配函数，根据语义距离动态分配语义间隔； The principle of similarity measure is: Firstly, the Euclidean distance between features is calculated as the semantic distance between samples. Secondly, a margin distributing function is designed to dynamically allocate margin in the basis of the semantic distances.
21400 最后，根据样本语义间隔计算误差并反向传播，使网络能够学习到样本间语义距离的差异，自动聚焦于难分样本，以充分学习难分样本的特征。 Finally, the difference of the sample semantic distance can be learned by calculating the loss and propagating it backwards so as to the network can automatically focus on the hard negative samples and more fully learn the characteristic of them.
21401 该网络在训练过程中采用多任务学习的方法，同时学习相似性度量和分类任务，以达到整体最优。 With a multi-task learning training method in the training stage, the similarity measure and classification can be learned jointly.
21402 在UCF101和HMDB51上的实验结果表明，与已有方法相比，提出的方法能有效提高视频分类精度。 Experimental results on UCF101 and HMDB51 show that the proposed method can effectively improve the classification precision.
21403 针对现有的基于用户轨迹的跨社交网络用户身份识别算法未考虑用户轨迹中的位置访问顺序特征的缺点，该文提出一种基于Paragraph2vec的跨社交网络用户轨迹匹配算法(CDTraj2vec)。 The performance of trajectory based user identification is poor since the existing methods ignore the order feature of location sequence. To solve this problem, a Cross Domain Trajectory matching algorithm based on Paragraph2vec (CDTraj2vec) is proposed.
21404 首先将用户轨迹转化为易于处理的网格化表示，并按照一定的时间粒度、距离尺度对原始的用户轨迹进行划分，使用户轨迹中的位置访问顺序特征易于抽取； Firstly, the user trajectory is transformed to the grid representation which is easy to handle. The PV-DM model in the Paragraph2vec algorithm is utilized for extracting order feature of location sequence in trajectory. Then the original user trajectories are divided by a certain time size and distance scale to construct a training sample suitable for training PV-DM model.