深度学习识别光网络单元故障的设计与应用

为解决依赖装维上门鉴别光网络单元故障带来的不便,可以从机器视觉入手实现自动化故障识别。近年,ImageNet挑战赛的成功推动了物体识别技术的跨越式发展,特别是基于卷积的深度学习技术在视觉识别方面已经达到人类水平,为光网络单元故障的自动识别提供了技术基础。文章对识别光网络单元的工作状态进行了研究,将设备工作状态分为7个场景,提出了利用手机APP采集图片识别故障的解决方案并投入了实际生产;重点阐述了深度学习模块的设计与实现,提出一种通过算法整合的方式综合运用物体检测和图像分类算法,分3阶段逐步求精,解决了图片过滤,光网络单元型号和状态识别等问题,实现了基于计算机视觉自动识别光网络单元故障。从数据上看产品的端到端准确率超过84%,识别速度达到10 FPS,月均提供服务超过1万人次,在减少用户等待的同时节约了人力资源。     

基于深度学习的视频跟踪研究进展综述

近年来深度学习迅猛发展，颠覆了语音识别、图像分类、文本理解等领域的算法设计思路。深度学习因其具备强大的特征提取能力，在图像识别领域的成绩尤为突出。然而深度学习与视频监控领域的结合并不多，由于深度模型具有多层网络结构，算法复杂度大，训练和更新模型时比较耗时，很难满足实时性要求。回顾了深度学习的发展史，介绍了最近10年来国内外深度学习主要模型，论述了基于深度学习的目标跟踪算法，指出了各算法的优缺点，最后对当前该领域存在的问题和发展前景进行了总结和展望。     

Dhh1 is a member of the SESA network

The correct separation of chromosomes during mitosis is necessary to prevent genetic instability and aneuploidy, which are responsible for cancer and other diseases, and it depends on proper centrosome duplication. In a recent study, we found that Smy2 can suppress the essential role of Mps2 in the insertion of yeast centrosome into the nuclear membrane by interacting with Eap1, Scp160, and Asc1 and designated this network as SESA (S my2, E ap1, S cp160, A sc1). Detailed analysis showed that the SESA network is part of a mechanism which regulates translation of POM34 mRNA. Thus, SESA is a system that suppresses spindle pole body duplication defects by repressing the translation of POM34 mRNA. In this study, we performed a genome-wide screening in order to identify new members of the SESA network and confirmed Dhh1 as a putative member. Dhh1 is a cytoplasmic DEAD-box helicase known to regulate translation. Therefore, we hypothesized that Dhh1 is responsible for the highly selective inhibition of POM34 mRNA by SESA.     

Automated classification of simulated wind field patterns from multiphysics ensemble forecasts

In this study, we have proposed an automated classification approach to identify meaningful patterns in wind field data. Utilizing an extensive simulated wind database, we have demonstrated that the proposed approach can identify low‐level jets, near‐uniform profiles, and other patterns in a reliable manner. We have studied the dependence of these wind profile patterns on locations (eg, offshore vs onshore), seasons, and diurnal cycles. Furthermore, we have found that the probability distributions of some of the patterns depend on the underlying planetary boundary layer schemes in a significant way. The future potential of the proposed approach in wind resource assessment and, more generally, in mesoscale model parameterization improvement is touched upon in this paper.     

视频侦查中人像智能分析应用及算法优化

人像智能分析指的是对视频或录像中的人像进行结构化和可视化分析,对目标人物进行性别、年龄、发型等特征的智能识别,这项技术在视频侦查中有极高的应用价值。人像识别早期的算法是通过人工提取特征,通过学习低级视觉特征来针对不同属性进行分类学习,这种基于传统方法的模型表现常常不尽如人意。在计算机视觉领域,通过海量图像数据学习的神经网络比传统方法有更丰富的信息量和特征可以被提取。文章尝试通过深度学习技术训练神经网络模型对行人进行检测和识别,对于衣着不同的行人进行智能识别,具有更好的鲁棒性,提升了视频人像识别的准确率,拓展了人工智能技术在身份识别领域的应用。     

基于特征融合的中文简历解析方法研究

针对基于规则和统计的传统中文简历解析方法效率低、成本高、泛化能力差的缺点，提出一种基于特征融合的中文简历解析方法，即级联Word2Vec生成的词向量和用BLSTM（Bidirectional Long Short-Term Memory）建模字序列生成的词向量，然后再结合BLSTM和CRF（Conditional Random Fields）对中文简历进行解析（BLSTM-CRF）。为了提高中文简历解析的效率，级联包含字序列信息的词向量和用Word2Vec生成的词向量，融合成一个新的词向量表示；再由BLSTM强大的学习能力融合词的上下文信息，输出所有可能标签序列的分值给CRF层；再由CRF引入标签之间约束关系求解最优序列。利用梯度下降算法训练神经网络，使用预先训练的词向量和Dropout优化神经网络，最终完成对中文简历的解析工作。实验结果表明，所提的特征融合方法优于传统的简历解析方法。     

Detection of correlation characteristics between financial time series based on multi-resolution analysis

Interactions between financial time series are complex and changeable in both time and frequency domains. To reveal the evolution characteristics of the time-varying relations between bivariate time series from a multi-resolution perspective, this study introduces an approach combining wavelet analysis and complex networks. In addition, to reduce the influence the phase lag between the time series has on the correlations, we propose dynamic time-warping (DTW) correlation coefficients to reflect the correlation degree between bivariate time series. Unlike previous studies that symbolized the time series only based on the correlation strength, the second-level symbol is set according to the correlation length during the coarse-graining process. This study presents a novel method to analyze bivariate time series and provides more information for investors and decision makers when investing in the stock market. We choose the closing prices of two stocks in China’s market as the sample and explore the evolutionary behavior of correlation modes from different resolutions. Furthermore, we perform experiments to discover the critical correlation modes between the bull market and the bear market on the high-resolution scale, the clustering effect during the financial crisis on the middle-resolution scale, and the potential pseudo period on the low-resolution scale. The experimental results exactly match reality, which provides powerful evidence to prove that our method is effective in financial time series analysis.     

Passive target detection and tracking from electromagnetic field measurements

This paper presents a novel approach to the localization of moving targets in a complex environment based on the measurement of the perturbations induced by the target presence on an independently‐generated time‐varying electromagnetic field. Field perturbations are measured via a set of sensors deployed over the domain of interest and used to detect and track a possible target by resorting to a particle Bernoulli filter (PBF). To comply with real‐time operation, the PBF works along with an artificial neural network (ANN) model of the environment trained offline via finite elements (FEs). The performance of the proposed algorithm is assessed via simulation experiments.     

基于注意力机制与文本信息的用户关系抽取

随着社交媒体的发展,用户之间的关系网络对于社交媒体的分析有很大的帮助。因此,该文主要研究用户好友关系检测。以往的关于用户好友关系抽取的研究主要基于社交媒体上的结构化信息,比如其他好友关系,用户的不同属性等。但是,很多时候用户本身并没有大量的好友信息存在,同时也不一定有很多确定的属性。因此,我们希望基于用户发表的文本信息来对用户关系进行预测。不同于以往的潜在好友推荐算法,该文提出了一种基于注意力机制以及长短时记忆网络(long short-term memory,LSTM)的好友关系预测模型,将好友之间的评论分开处理,通过分析用户之间的评论来判断是否具备一定的好友关系。该模型将好友双方信息拼接后的结果作为输入,并将注意力机制应用于LSTM的输出。实验表明,用户之间的评论对于好友关系预测确实有较大的实际意义,该文提出的模型较之于多个基准系统的效果,取得了明显的提升。在不加入任何其它非文本特征的情况下,实验结果的准确率达到了77%。     

Food recognition using neural network classifier and multiple hypotheses image segmentation

ABSTRACT

This paper proposes the multiple-hypotheses image segmentation and feed-forward neural network classifier for food recognition to improve the performance. Initially, the food or meal image is given as input. Then, the segmentation is applied to identify the regions, where a particular food item is located using salient region detection, multi-scale segmentation, and fast rejection. Then, the features of every food item are extracted by the global feature and local feature extraction. After the features are obtained, the classification is performed for each segmented region using a feed-forward neural network model. Finally, the calorie value is computed with the aid of (i) food volume and (ii) calorie and nutrition measure based on mass value. The experimental results and performance evaluation are validated. The outcome of the proposed method attains 0.947 for Macro Average Accuracy (MAA) and 0.959 for Standard Accuracy (SA), which provides better classification performance.