ID 原文 译文
56298 针对上述问题,本文提出了一种基于对抗学习和语义相似度的跨媒体搜索方法,实现了文本和图像之间的相互匹配、排序和搜索. Focusing on these problems, in this paper, we propose a cross?media search method based on adversarial learning and semantic similarity to accomplish matching and sortingbetween text and image.
56299 该方法使用对抗学习方法框架构建训练特征映射网络和模态判别网络,其中特征映射网络使用多维语义分布向量将不同模态的数据映射到同一语义空间中,使得相同语义下的不同模态数据在该空间距离小,不同语义下相同模态数据距离大. This method mainly adopts adversarial learning to train a feature projecting networkand a modal specific network. The feature mapping network working as a generator converts raw data fromdifferent modalities into a common semantic space based on the corresponding semantic distributions, throughwhich the distance of data from the different modalities is small under the same semantic space while the distanceof data from the same modal under different semantics is large.
56300 使用语义分布及相似度作为特征映射网训练依据,模态判别网络负责判定空间中不同数据的模态. The generator uses semantic distributions andsimilarities as the basis for training the feature mapping network. The discriminator aims at distinguishing themodality of the data from different media projected into the common semantic space.
56301 基于对抗学习交替训练两个网络,使得特征映射网络得到的数据和原数据语义一致,并消除模态特性,最终在同一空间内使用相似度来排序并得到搜索结果. The designed method withtwo networks is alternately trained until convergence when the features obtained by the feature mapping networkare consistent with the original data semantics, and the modal feature is eliminated. Finally, the search resultsare obtained by similarity ranking based on the raw data from the common semantic space.
56302 实验结果表明本文提出的方法在文本和图像的相互搜索的map值比同类方法高,并验证了该方法在社交网络安全话题数据上的有效性. The efficiency resultsshow that the evaluation value of text and image search is higher than other similar methods. The efficiency ofthe method is verified on social network data on security topics
56303 光流 (optical flow) 为同一对象在视频中运动到下一帧的移动量. Optical flow plays a key role in recording the movement of an object in videos to its next frame.
56304 从视频中估计光流已广泛应用于各类移动智能系统, 如运动估计和机器人导航. Estimating optical flow from a video has been widely applied in many mobile cognitive systems such as motionestimation and robot navigation.
56305 最近的研究表明, 卷积神经网络 (convolutional neuralnetwork, CNN) 能提供可靠的光流估计结果.然而, 现有的硬件加速器无法支持面向光流估计的 CNN复杂计算. Recent works have demonstrated that optical flow estimation can be successfullysolved by emerging convolutional neural networks (CNNs). Existing dedicated accelerators, however, cannotsupport the complex computation of optical flow oriented CNNs.
56306 具体而言, 这些类型的 CNN 不仅包括常规的卷积 (convolution) 和反卷积 (deconvolution)运算, 还包括双线性插值 (bilinear interpolation) 和/或关联 (correlation) 运算. Specifically, these CNNs are composed ofnot only the conventional convolution and deconvolution operations, but also the bilinear interpolation and/orcorrelation operations, which exploit the correspondence of two consecutive image frames.
56307 双线性插值和关联操作主要探索两个连续图像帧之间的关联关系. To address this problem,we propose an integrated accelerating solution called Swan-AOE aiming to tackle optical flow oriented CNNs, i. e. ,by supporting convolution, deconvolution, bilinear interpolation, and correlation layers.