基于深度CRF模型的图像语义分割方法

从图像中提取多种特征向量堆叠为一个高维特征向量用于图像语义分割,会导致部分特征向量的分类能力减弱或丢失。针对此问题,提出了一种结合深度卷积神经网络AlexNet和条件随机场的图像语义分割方法。利用预训练好的AlexNet模型提取图像特征,再通过条件随机场对多特征及上下文信息的有效利用来实现图像的语义分割。与利用传统经典特征的方法进行对比,实验结果表明:在利用AlexNet模型提取特征进行图像语义分割时,Conv5层为最有效的特征提取层,在Stanford background和Weizmann horse数据集下的识别准确率分别为81.0%和91.7%,均高于其他2种对比方法,说明AlexNet可以提取更有效的特征,得到更高的语义分割精度。     

低维显式语义空间下的语义关联度计算方法

语义关联度计算是数据科学中的一个关键性基础问题,在信息检索及自然语言处理等方面有着广泛的应用.针对ESA (Explicit Semantic Analysis)算法存在的局限性,提出一种显式语义特征选择算法,并构建低维语义空间.在此基础上,根据特征概念在Wikipedia中的映射信息,提出一种低维显式语义空间下的语义关联度计算方法.该方法解决了ESA算法在后续语义关联度计算过程中,因高维稀疏空间导致计算效果不够准确的问题.实验结果表明,与当前其他方法相比,该方法的计算结果在皮尔逊相关系数(P)及斯皮尔曼相关系数(S)上与人们的认知判断之间具有更好的一致性.     

足球比赛场景中交互行为分析方法研究

目前的视频处理研究主要关注足球视频的基础方面的问题, 而在视频语义方面的工作尚不完善。 为此, 对足球比赛视频中存在的交互行为进行分析, 研究了足球比赛视频语义理解的关键问题, 为实现从一般比赛场景去理解和分析比赛进行了有益的探索。 首先进行场景分析, 利用主颜色提取视频中场地区域并标定场地线; 其次, 检测场景中的球员并判定其队属, 同时利用 Kalman 滤波实现对足球目标的跟踪; 最后, 判定球队的进攻方向的目的语义, 同时获取足球真实轨迹, 进行轨迹分割, 识别球员与足球的交互行为, 并在此基础上结合足球领域的知识进行进攻策略的分析。 实验结果表明, 该算法能很好地解决视频分析各阶段的问题, 能有效分析视频中存在的交互行为, 并从足球领域的角度实现进攻策略分析,进攻模式识别结果查准率达到 71% 以上, 查全率达到 85% 以上, 有益于计算机视觉的应用研究和足球视频科学分析。     

Semantic image annotation based on GMM and random walk model

Automatic image annotation has been an active topic of research in computer vision and patternrecognition for decades.A two stage automatic image annotation method based on Gaussian mixturemodel (GMM) and random walk model (abbreviated as GMM-RW) is presented.To start with,GMM fitted by the rival penalized expectation maximization (RPEM) algorithm is employed to estimatethe posterior probabilities of each annotation keyword.Subsequently, a random walk processover the constructed label similarity graph is implemented to further mine the potential correlations ofthe candidate annotations so as to capture the refining results, which plays a crucial role in semanticbased image retrieval.The contributions exhibited in this work are multifold.First, GMM is exploitedto capture the initial semantic annotations, especially the RPEM algorithm is utilized to train themodel that can determine the number of components in GMM automatically.Second, a label similaritygraph is constructed by a weighted linear combination of label similarity and visual similarity ofimages associated with the corresponding labels, which is able to avoid the phenomena of polysemyand synonym efficiently during the image annotation process.Third, the random walk is implementedover the constructed label graph to further refine the candidate set of annotations generated byGMM.Conducted experiments on the standard Corel5k demonstrate that GMM-RW is significantlymore effective than several state-of-the-arts regarding their effectiveness and efficiency in the task of automatic image annotation.     

2D difference gel electrophoresis reference map of a Fusarium graminearum nivalenol producing strain

Fusarium graminearum is widely studied as a model for toxin production among plant pathogenic fungi. A 2D DIGE reference map for the nivalenol‐producing strain 453 was established. Based on a whole protein extract, all reproducible spots were systematically picked and analyzed by MALDI‐TOF/TOF, leading to the identification of 1102 protein species. The obtained map contributes to the annotation of the genome by identifying previously nondescribed hypothetical proteins and will serve as a reference for future studies aiming at deciphering F. graminearum biology and chemotype diversity.     

TargetATPsite: A template‐free method for ATP‐binding sites prediction with residue evolution image sparse representation and classifier ensemble

Understanding the interactions between proteins and ligands is critical for protein function annotations and drug discovery. We report a new sequence‐based template‐free predictor (TargetATPsite) to identify the Adenosine‐5′‐triphosphate (ATP) binding sites with machine‐learning approaches. Two steps are implemented in TargetATPsite: binding residues and pockets predictions, respectively. To predict the binding residues, a novel image sparse representation technique is proposed to encode residue evolution information treated as the input features. An ensemble classifier constructed based on support vector machines (SVM) from multiple random under‐samplings is used as the prediction model, which is effective for dealing with imbalance phenomenon between the positive and negative training samples. Compared with the existing ATP‐specific sequence‐based predictors, TargetATPsite is featured by the second step of possessing the capability of further identifying the binding pockets from the predicted binding residues through a spatial clustering algorithm. Experimental results on three benchmark datasets demonstrate the efficacy of TargetATPsite. © 2013 Wiley Periodicals, Inc.     

基于问句语义表征的中文问句相似度计算方法

提出了一种中文问句语义相似度计算的新方法.该方法分为两步:第一步采用基于问句句型模板规则匹配的方法提取问句语义表征;第二步根据问句语义表征计算问句语义相似度.采用该方法开发了一个面向常问问题集(FAQ)的问答系统.实验结果表明,采用该方法获得的相似度计算的准确率约为85%.     

Evaluation of ion mobility in capillary electrophoresis coupled to mass spectrometry for the identification in metabolomics

Currently, feature annotation remains one of the main challenges in untargeted metabolomics. In this context, the information provided by high-resolution mass spectrometry (HRMS) in addition to accurate mass can improve the quality of metabolite annotation, and MS/MS fragmentation patterns are widely used. Accurate mass and a separation index, such as retention time or effective mobility (μ_eff), in chromatographic and electrophoretic approaches, respectively, must be used for unequivocal metabolite identification. The possibility of measuring collision cross-section (CCS) values by using ion mobility (IM) is becoming increasingly popular in metabolomic studies thanks to the new generation of IM mass spectrometers. Based on their similar separation mechanisms involving electric field and the size of the compounds, the complementarity of ^DTCCS_N2 and μ_eff needs to be evaluated. In this study, a comparison of ^DTCCS_N2 and μ_eff was achieved in the context of feature identification ability in untargeted metabolomics by capillary zone electrophoresis (CZE) coupled with HRMS. This study confirms the high correlation of ^DTCCS_N2 with the mass of the studied metabolites as well as the orthogonality between accurate mass and μ_eff, making this combination particularly interesting for the identification of several endogenous metabolites. The use of IM-MS remains of great interest for facilitating the annotation of neutral metabolites present in the electroosmotic flow (EOF) that are poorly or not separated by CZE.     

Cheminformatics in MS-based environmental exposomics: Current achievements and future directions

Compound annotation using MS/MS data is the major bottleneck in interpretation of mass spectrometry data during non-targeted screening and suspect screening exposomics studies. Apart from compound identification using available databases or mass spectral libraries, the true challenge comes when completely new compounds have to be identified. Along with recent advances in MS instrumentation that set grounds to a new revolutionary age in environmental exposomics, a multitude of cheminformatics annotation approaches has been developed. Herein, we review the basic principles of the cutting-edge cheminformatics MS-based approaches employed in eco-exposome annotation.We give a solid background discussing the eco-exposome concept in relation to the advances in MS instrumentation, and define the three crucial cheminformatics tasks used in the eco-exposome annotation: molecular formula assignment, compound prioritization and compound annotation. The basic principles of compound annotation are discussed, which are based on three approaches of utilizing structural information inherent to MS data. These involve direct, indirect and joint annotation approaches. We assess their performance through the ability to annotate eco-exposome constituents. We discuss future perspectives and give directions to new annotation strategies and performance evaluation protocols aiming to solve current issues hampering the incorporation of cheminformatics annotation approaches in regular eco-exposome annotation workflows.     

基于维基语义的多文档文摘研究

多文档文摘作为自然语言处理领域的重要技术之一,能从不同角度辅助用户实现高效的信息获取.由于文档集合内的内容往往来自不同的信息源,文本之间通常存在丰富而复杂的语义关系.常用的基于词的文档表示法,难以为文摘的语义分析过程提供充足而准确的数据信息.为此,我们提出使用维基百科——当今世界最大的在线概念语料库——为多文档文摘的提取提供语义支持.一方面,我们通过提取文档中的维基概念,生成准确一致的句子表示形式.另一方面,在计算句子特征时,我们利用维基词条的首段指导机器文摘的提取.我们首先通过计算概念在维基中的全局相关性和当前文档集内的局部相关性,获取概念的权重.然后在维基概念表示的基础上,为文档中的句子提取多种基于维基的特征,并最后用于文摘生成.在实验中,我们依次用各个维基特征独立生成文摘,并使用ROUGE(Recall-Oriented Understudy for Gisting Evaluation,面向召回率的要点评估)指标评价文摘质量.通过比较,实验验证了维基词条首段能较好的提升文摘质量.