基于空间域和频率域特征融合的场景文本识别

对于小样本语言无关场景的文本识别,现有的方法往往面临鲁棒性低和泛化能力差的问题。针对这一问题,一方面,在特征提取阶段,提出了基于空间域和频率域特征融合的双流网络结构,其包含一个提取空间域特征的深度残差卷积网络分支,以及提取频率域特征的一维快速傅里叶变换和浅层神经网络分支,接着使用通道注意力机制融合这两种特征。另一方面,在序列建模阶段,针对语言无关场景的特点,提出一种多尺度一维卷积模块用来代替双向长短期记忆网络。然后结合现有的TPS矫正模块和CTC解码器搭建完整模型。训练过程中采用了迁移学习的方法,先在大型英文数据集上进行预训练,后在目标数据集上进行微调。在文中整理的两个小样本语言无关数据集上的实验结果表明,所提模型在准确率上优于现有的模型,验证了其在该场景下的具有较高的鲁棒性和泛化能力;此外,在语言相关场景的5个基准数据集上的相关实验(不用微调)表明,使用文中所述特征提取模块的方法优于对比的基线方法,证明了所提出的双流特征融合网络的有效性和通用性。

电力城域远端分支机构网络接入方案探讨

针对目前电力城域远端分支机构的布点及电缆布放特点,对其接入电力信息中心网络的方案进行分析,总结出千兆以太网无源光网络(EPON)在此类接入应用的优越性及设计要点。

10 kV配电网含分支混合线路故障行波定位研究

考虑配电网实际结构,建立配电网常见的架空线—电缆混合线路模型,设置分支点和分支线路,模拟分界开关和环网柜及其出线情况,提出分支域的概念。基于此提出一套故障定位方法,首先比较分支域外和域内故障时各分支线路末端与支路分支点初始行波到达时差的不同,以确定故障是否发生于分支线路;其次将实际故障时混合主干线路初始行波到达始末端的时差值与分支点和线缆连接点故障时相比较,并结合分支线路故障判据以确定主干线路的故障区段;最后提出简单的单双端行波组合测距方法,确定具体故障位置。PSCAD仿真结果证明,该套定位方法能较为准确地选出故障线段,并确定故障位置。

基于双流网络结构的深度伪造人脸的检测方法

深度伪造技术(Deepfake)是一种基于生成对抗网络(Generative Adversarial Networks,GAN)的深度网络模型,可以利用源和目标人脸生成高度逼真且难以鉴别的人脸视频。如果不法分子借此技术制造虚假视频并在互联网上传播谣言,将会侵犯个人肖像权,造成不良的社会影响,甚至引发严重的司法纠纷。面对深度伪造技术带来的严重威胁,国内外众多研究机构高度关注深度伪造检测技术的研究并提出了若干检测方法。现有的检测方法在高质量视频上可以取得良好的检测效果,然而日常应用中的视频通常会通过社交软件从而被压缩为低质量视频,在此类低质量数据集中,现有的大多数伪造人脸检测方法的准确率有着明显的下降,并且现有方法在跨库情况下的检测性能也不够理想。文中针对现有工作的局限性,提出了一种注意力机制下基于Xception模型的双流网络结构。该网络结构中包含了使用多重注意力机制的RGB分支,以及用于捕捉低质量视频伪影效应的频率域分支。通过研究发现,真实图像与伪造图像之间的微小差别更多地集中在局部位置,因此多重注意力机制下的RGB分支将使得模型关注人脸的不同区域,并在注意力图的指导下得到由低层纹理特征及高层语义特征聚合的全局特征。频率域分支引入离散余弦变换作为频域变换手段,为图像提供与RGB分支互补的特征表示,此分支能够反映细微的伪造痕迹或者压缩误差。为了验证该网络结构的有效性,所提算法在FaceForensics++,Celeb-DF以及DFDC 3个公开数据集上进行了大量对比实验。实验结果表明,所提算法在低质量视频集上的性能优于现有的检测算法,并且所提模型在跨库场景下具有更好的检测性能,即验证了文中提出的注意力机制下的RGB和频率域双流特征的结合可以提高检测模型在低质量视频集及跨库情形下的鲁棒性。

Distribution area demarcation for logistic centers under geographical information system decision support system

Demarcating distribution area of goods is often guided by the rule of thumb by business proprietors. However, this method seems to be unsuitable when the demand points increase to a certain large extent. The present work attempted to convert the problem of distribution area demarcation into a localized problem of warehouseing and networking, and tried to establish district-based planning mode based on location based heuristic (LBH). Two methods were used in this study: 1) the manual method to construct the mathematical model and conduct simulation; 2) the automatic method using TransCAD software of geographical information system (GIS) for simulation. By comparing the effects of the two methods, the research provides theoretical support for business proprietors to demarcate the distribution area rationally with the application of GIS system. The results show that GIS has very good graphics construction function to replace complex text, and the automatic demarcating mode with human-machine interaction provides a good business decision-making support.

Rolling bearing performance degradation evaluation by VMD and embedding selection-based NPE

In order to improve the incipient fault sensitivity and stability of degradation index in the rolling bearing performance degradation evaluation process,an embedding selection-based neighborhood preserving embedding(ESNPE)method is proposed.Firstly,the acquired vibration signals are decomposed by variational mode decomposition(VMD),and the singular value and relative energy of each intrinsic mode function(IMF)are extracted to form a high-dimensional feature set.Then,the NPE manifold learning method is used to extract the embedded features in the feature space.Considering the problem that useful embedding information is easily suppressed in NPE,an embedding selection strategy is built based on the Spearman correlation coefficient.The effectiveness of embeddings is measured by the coefficient absolute value,and useful embeddings are preserved in the early stage of bearing degradation by using the first-order difference method.Finally,the degradation index is established using the support vector data description(SVDD)model and bearing performance degradation evaluation is achieved.The proposed method was tested with the whole life experiment data of a rolling bearing,and the result was compared with the feature extraction methods of traditional principal component analysis(PCA)and NPE.The results show that the proposed method is superior in improving the incipient fault sensitivity and stability of the degradation index.

C_(p,q)^(s+α)-forms Solution of _b-equ ation and Its L_(p,q)~s Estimates

In this paper, by using holomorphic support f unction of strictly pseudoconvex domain on Stein manifolds and the kernel define d by DEMAILY J P and Laurent Thiebaut, we construct two integral operators T q and S q which are both belong to C s+α p,q-1 (D) and ob tain integral representation of the solution of (p,q)-form b-equation on the boundary of pseudoconvex domain in Stein manifolds and the L s p,q extimates for the solution.

Multiple walking human recognition based on radar micro-Doppler signatures

The recognition of human movements based on radar m-D(micro-Doppler) signatures attracts great interest in the field of radar research on automatic target recognition. Because there are multiple frequency components overlapping seriously in the radar echoes from walking humans, it is a very difficult work to recognize walking humans based on radar echoes. In this paper, a recognition method of walking humans based on radar m-D signatures is proposed. In this method, the m-D spectrum is generated by generalized S transform first,and then the entropy segmentation is used to segment the interesting region from the original spectrum. Next,the m-D features are extracted from the m-D region. Lastly, the support vector machine is used to recognize different walking human targets. The simulation experiments considering two factors of height and velocity are also conducted to test the performance of this proposed method.