基于多流架构与长短时记忆网络的组群行为识别方法研究

提出一种基于多流架构与长短时记忆网络的上下文建模框架,旨在克服组群行为识别的两个难点,其一对复杂场景中多视觉线索进行信息融合;其二对情景人物进行建模,以获得长视频上下文关系.并且,对基于全局信息和基于局部信息的识别结果进行决策融合,判定最终组群行为属性.该算法在CAD1和CAD2上分别取得93.2%和95.7%平均识别率.     

基于可变形模板匹配的变形字体识别

针对变形字体的识别问题,提出了一种基于形状上下文和模糊推理相结合的度量机制的变形字体识别方法。首先,在分析现行字体识别方法的基础上,提出了薄板样条曲线TPS与模糊推理相结合的相似性距离度量机制的总体思路;其次,描述了基于形状上下文的变形字体图像的特征提取过程;而后讨论了薄板样条曲线TPS与模糊推理相结合的相似性距离度量机制,并将其应用在变形字体的识别算法中。实验结果表明所提出的方法是有效的。     

一种本体和上下文知识集成化的数据挖掘方法

在数据挖掘中使用本体和上下文知识能够将普遍的知识和特定的知识引入数据挖掘的决策因素中,是增进数据挖掘准确性的有效手段,同时也是数据挖掘领域研究的热点和难点之一.针对该问题,首先探讨了本体与上下文知识的集成化表示方法,包括上下文知识分类方法、如何在本体描述方法上扩展上下文知识及上下文知识转化方法.其次,以层次化结构的本体与上下文知识为例,构建了一个依据于本体和上下文知识集成的归纳学习算法并验证了该算法的有效性和准确性.     

基于时空上下文和随机森林的人眼跟踪定位算法研究

为了得到人眼跟踪过程中更好的鲁棒性和实时性以及跟踪精度,提出一种基于自适应增强分类算法(AdaBoost)、随机森林(RF)和时空上下文(STC)的重定位跟踪算法。该算法结构分为3层,分别为AdaBoost人脸检测、STC人脸跟踪和RF人眼定位。首先,利用AdaBoost在第一帧识别出人脸,从而提取出人脸窗口。接着,使用时空上下文跟踪算法进行人脸跟踪。然后,联合定向梯度直方图(HOG)算法进行相似度判断,以达到目标丢失后继续跟踪的目的。最后,采用随机森林算法进行人眼定位。实验结果表明,与传统的随机森林人眼跟踪算法相比,该算法在跟踪速度达到原方法的2倍左右,并在跟踪精度和鲁棒性上和原算法相同。基本满足在裸眼3D显示时人脸跟踪和人眼定位的精度高、实时性快、鲁棒性好的要求。     

基于双向LSTM的误植域名滥用检测方法

当前,误植域名检测主要以计算域名对之间的编辑距离为基础,未能充分挖掘域名的上下文信息,且对短域名的检测易产生大量的假阳性结果。采集域名相关信息进行判定虽然有助于提高检测效果,却会引入较大的额外开销.本文采用了基于域名字符串的轻量级检测策略,并引入双向长短时记忆模型（LSTM,Long Short-Term Memory）来充分利用域名上下文,提升检测效果.本文还设计了面向域名的局部敏感哈希函数,以提高在大规模域名集合上进行误植域名检测的速度.在大量真实数据集上的实验结果表明,本文的工作改进了基于编辑距离检测方法的不足,能够有效地进行误植域名滥用检测.     

时序行为提名的上下文信息融合方法

在针对视频的人体活动定位和识别领域中,现有的时序行为提名方法无法很好地解决行为特征长期依赖性而导致提名召回率较低.针对此问题,提出了一种上下文信息融合的时序行为提名方法.该方法首先采用三维卷积网络提取视频单元的时空特征,然后采用双向门控循环网络构建上下文关系预测出时序行为区间.针对门控循环单元(GRU)存在参数较多和梯...     

可解释深度知识追踪模型

知识追踪任务通过建模用户的习题作答序列跟踪其认知状态,进而预测其下一时刻的答题情况,实现对用户知识掌握程度的智能评估.当前知识追踪方法多针对知识点建模,忽略了习题信息建模与用户个性化表征,并且对于预测结果缺乏可解释性.针对以上问题,提出了一个可解释的深度知识追踪框架.首先引入习题的上下文信息挖掘习题与知识点间的隐含关系,得到更有表征能力的习题与知识点表示,缓解数据稀疏问题.接着建模用户答题序列获得其当前知识状态,并以此学习个性化注意力,进而得到当前习题基于用户知识状态的个性化表示.最后,对于预测结果,依据个性化注意力选择一条推理路径作为其解释.相较于现有方法,所提模型不仅取得了更好的预测结果,还能为预测结果提供推理路径层面的解释,体现了其优越性.     

Speaking rate from proximal and distal contexts is used during word segmentation.

A series of eye-tracking and categorization experiments investigated the use of speaking-rate information in the segmentation of Dutch ambiguous-word sequences. Juncture phonemes with ambiguous durations (e.g., [s] in 'eens (s)peer,' “once (s)pear,” [t] in 'nooit (t)rap,' “never staircase/quick”) were perceived as longer and hence more often as word-initial when following a fast than a slow context sentence. Listeners used speaking-rate information as soon as it became available. Rate information from a context proximal to the juncture phoneme and from a more distal context was used during on-line word recognition, as reflected in listeners' eye movements. Stronger effects of distal context, however, were observed in the categorization task, which measures the off-line results of the word-recognition process. In categorization, the amount of rate context had the greatest influence on the use of rate information, but in eye tracking, the rate information's proximal location was the most important. These findings constrain accounts of how speaking rate modulates the interpretation of durational cues during word recognition by suggesting that rate estimates are used to evaluate upcoming phonetic information continuously during prelexical speech processing. (PsycINFO Database Record (c) 2011 APA, all rights reserved)     

A guiding framework and approach for implementation research in substance use disorders treatment.

This paper introduces readers to the concepts of implementation science, implementation theory, and implementation frameworks and models. A wide range of models has been published in the literature related to implementation. The paper will present an overview of the Consolidated Framework for Implementation Research (CFIR), which is a comprehensive typology that unifies and consolidates the array of constructs that influence implementation from the perspective of these models. The CFIR is then used to evaluate implementation models used in studies of substance use disorder (SUD) treatments. Implementation research is scarce, with few prospective studies of theory-driven implementation. We assert that future research in SUD needs to meet three overarching objectives to promote wider implementation of evidence-based practices: (a) differentiation of core versus adaptable components of evidence-based interventions need; (b) development of methods to design implementation strategies, effectively adapted to the broad context; and (c) design and testing of predictive models to assess likelihood of effective implementation and prospects for sustainability while taking into account salient contextual factors. A recommended strategy for accomplishing these objectives is described. (PsycINFO Database Record (c) 2011 APA, all rights reserved)