基于深度学习CNN模型的图像情感特征抽取研究

[目的/意义]以用户情感为线索的图像检索已成为机器学习研究的热点，但图像情感特征标注的语料数据多来源于对图像低层特征的抽取，从而导致图像检索过程单一化和程式化。本文提出了一种基于深度学习的图像情感特征抽取的算法，将图像底层特征融合到图像的高层情感语义当中，为实现图像的情感语义检索提供了参考。[方法/过程]利用改进的卷积网络模型，将数据集图像的颜色、纹理作为输入，经多层运算自动提取图像的情感信息，并通过反向传播算法计算出改进后模型的情感检索准确率，构造出准确率较高且过拟合程度低的图像情感特征提取模型。[结果/结论]应用改进的卷积神经网络模型，实现了对图像情感特征的抽取，相较于原模型提升了10%的检索准确率。     

面向数字文旅的图像文本跨模态检索方法

[目的/意义]图像文本跨模态检索应用对最大化利用数字文旅资源具有重要意义。然而,数字文旅领域的图像文本跨模态检索方法面临长文本挑战、数据缺失、内存资源有限等问题。为此,我们提出了一种新的基于Transformers和MobileNet V3模型的数字文旅图像文本跨模态方法。[方法/过程]首先,提出了基于自注意力机制的双层多组Transformers模型从标题、正文和评论等文本中学习具有互补性的文本特征;其次,设计了FastR-CNN和MobileNet V3模型学习图像局部细粒度特征;最后,提出了多元线性回归方法在共享子空间补全缺失数据。构建以图搜文和以文搜图的双向三元损失函数学习模型参数。[结果/结论]在标准数据集Flickr30k、自建数据集CulTour-Sha和有数据缺失的数据集Flickr30k-1与CulTour-Sha-1上的大量实验结果表明,我们的方法在召回率、内存需求和计算速度等方面优于当前几种先进的跨模态检索方法。     

Material characterization,stratigraphy, textures,and painting techniques of the mural painting The Allegory of the Industrial Development of São Paulo by Fulvio Pennacchi

ABSTRACT

This work presents a study of samples of the mural painting The Allegory of the Industrial Development of São Paulo (Alegoria ao Desenvolvimento Industrial Paulista), made in the city of São Paulo, Brazil, by the artist Fulvio Pennacchi. Using transmitted light petrographic microscopy, scanning electron microscopy coupled with energy dispersive X-ray spectrometry, and X-ray fluorescence, the stratigraphy, the textures, the mineralogy, and the chemical compositions of the mural were established; therefore the painting techniques, pigments, and materials applied by the artist were identified. The mural was painted with the fresco technique and variations. This technique involves the use of materials of mineral origin, both in the support and for the pigments. About the artist’s relevance, Fulvio Pennacchi is considered an important muralist of his time, being one of the pioneers in the production of frescoes in Brazil. The mural, measuring approximately 11?×?6?m, painted in 1954, was removed from where it was originally made, in consequence of a fire that affected the place. Specialized engineers transported it, and today it is stored awaiting restoration. The data presented here can support the process of restoration of the mural, as well as contribute to the improvement of protocols for the analysis of works of art, and to the historical record of mural painting techniques over time.     

基于字词向量的多尺度卷积神经网络微博评论的情感分类模型及实验研究

[目的/意义]微博评论情感分类模型可以为相关舆情监管部门正确管控话题事件的发展状况和舆情提供一定的指导作用。[方法/过程]基于字词向量的多尺度卷积神经网络，运用多尺度卷积核改善微博评论中上下文信息有限的条件制约，构建基于字词向量的多尺度卷积神经网络微博评论情感分类模型；通过爬取"微博热搜整改"数据，对模型的可行性和优越性进行验证。[结果/结论]验证结果表明基于字词向量的多尺度卷积神经网络在微博舆情等上下文信息有限的短文本分类任务中表现良好。本文在理论层面为微博舆情情感分类提供了更为准确的情感分类理论模型及分类方法，在实践层面可以更好地指导舆情监管部门对舆情的情感倾向进行更好的引导和监管。     

面向敦煌壁画的移动视觉搜索模型研究

[目的/意义]保护敦煌文化遗产,为敦煌壁画提供移动视觉搜索服务,以助用户高效、便捷地获取敦煌壁画丰富知识资源。[方法/过程]构建基于BoW的图像底层特征匹配和基于主题标签的高层语义关联的移动视觉搜索模型,用SIFT提取图像局部特征,K-means生成有K个视觉单词组成的视觉词典,计算图像映射到视觉词典的TF-IDF向量,通过内积计算相似度匹配图像并排序;根据图像主题添加语义标签,提取最匹配图像的标签实现语义关联搜索;最后收集1200张敦煌壁画图片验证模型的有效性。[结果/结论]在视觉单词数为1000时BoW+SIFT的图像搜索耗时163ms,且图像大小在0.5-2.5倍范围的准确率在83.7%以上,可有效搜索主题语义关联图像。     

国外图像特征研究进展与启示

[目的/意义]调研分析国外图像特征研究的进展与发现,总结图像特征体系的特点以及相关用户使用行为,为改进和优化图像标引和检索提供参考借鉴。[方法/过程]系统回顾国外图像特征的相关研究,分析现有图像描述和著录规范、图像特征的分类与分层研究以及标引与检索情境下用户对图像特征的使用行为及影响因素等研究。[结果/结论]国外图像特征研究内容较为全面和系统,研究视角涵盖以图像为中心和以用户为中心,研究方法涉及日志研究和调查研究,研究对象有普通用户和专业人员,研究发现,图像类型、用户视觉经验、专业知识背景、任务类型、检索阶段等因素与图像特征使用行为有关。未来可在面向非专业用户的图像特征体系构建以及图像特征使用等方面开展更多研究。     

How do scholars and non-scholars participate in dataset dissemination on Twitter

Focusing on the dataset dissemination structure on Twitter, this study aims to investigate how users of two different identities, scholars and the public, participate in the dissemination process. We collected 2464 datasets from Altmetric.com and used social network analysis to plot the graphs. From a macroscopic viewpoint, most datasets were diffused by viral dissemination (structure II) and mixed dissemination (structure III), and the diffusion level was fundamentally one or two levels. Based on the topics clustering results of the datasets, the majority were about open access, research data, and Altmetrics, as well as astronomy, biology, medicine, and environmental engineering. The dataset dissemination structure shared a little relationship with the research topic. From the microscopic viewpoint of parent nodes and child nodes, during the dataset dissemination, there were only marginally more Twitter users with scholar status than non-scholar ones, suggesting that compared with traditional academic accomplishments such as journal papers. However, the dataset seems to be more professional and targeted; significant audience beyond academics are also involved. During disseminating datasets on Twitter, most tended to be diffused among users of the same identity. However, a few non-scholars played crucial roles, such as super users and intermediaries. Overall, a considerable part of tweets and tweets of parent nodes with the ability to spread is primarily the tweets commented simultaneously forwarded (type II) are posted at the same time commented. Hence, this study underlines the significance of research data-sharing and social media's role in public participation in science.     

学术文本词汇功能识别--在关键词自动抽取中的应用

传统的关键词自动抽取常以候选词的出现频次、位置等非语义信息构建特征,并未考虑关键词在学术文献中承担的特定语义角色,即词汇功能。通过对现有数据统计,本文发现作者标注关键词中约有67.99%是研究问题或研究方法词。因此,本文将关键词的词汇功能分为三类:“研究问题”“研究方法”和“其他”,在传统的词频特征以及位置特征基础上,融合词汇功能特征,使用计算机领域的学术文献基于分类和排序两种思想进行关键词抽取实验。实验结果表明,融合词汇功能后,关键词抽取效果得到明显提升。相较于基准实验,二分类模型的准确率Acc和F值分别相对提升24.63%和25.19%,达到了0.840和0.666;排序模型的MAP、NDCG@5和P@5分别相对提升168.32%、189.50%和148.30%,提升至0.813、0.828和0.447,证明了学术文献词汇功能特征在关键词自动抽取中具有重要作用。     

基于网络结构和文本内容的群体画像构建方法研究

[目的/意义] 在基于社会网络的用户画像研究中，针对传统用户建模难以处理复杂网络关系，群体构建多基于内容，以及群体相似度低或紧密性差的问题，提出基于网络结构和文本内容的群体画像构建方法。[方法/过程] 首先，采用卷积神经网络方法，融合网络结构和文本内容两方面特征将网络用户表示成空间向量，其次，在k-means算法基础上结合模块度计算方法，对空间向量进行聚类，然后，在爬取的中英文数据集上分别进行对比研究，最后，从中文数据集中选取1 000名重要性用户进行实例分析。[结果/结论] 实验结果表明，该方法的密度值比基于内容的方法平均增加0.105，熵值比基于结构（含基于结构和内容）的方法平均减少0.955，实例分析进一步说明文中方法的可行性。     

网络环境下艺术院校图书馆空间再造——以中国美术学院图书馆为例

[目的/意义]探讨与研究网络环境下艺术院校图书馆空间再造，以推进图书馆转型与发展。[方法/过程]采用文献研究法、网络调查法等方法，评述国内外图书馆空间再造的理论研究与实践，并以中国美术学院图书馆特色空间再造实践为例，阐述艺术院校图书馆如何在网络环境下对馆舍空间进行调整、改造和再利用。[结果/结论]网络环境下图书馆空间价值的日益凸显，成为图书馆服务的核心内容之一。中国美术学院图书馆通过学术研讨空间、艺术展览空间、书画临摹空间等空间再造项目的实践，拓展图书馆服务功能，提升空间利用效率。     

A decision support system: Automated crime report analysis and classification for e-government

This paper investigates how text analysis and classification techniques can be used to enhance e-government, typically law enforcement agencies' efficiency and effectiveness by analyzing text reports automatically and provide timely supporting information to decision makers. With an increasing number of anonymous crime reports being filed and digitized, it is generally difficult for crime analysts to process and analyze crime reports efficiently. Complicating the problem is that the information has not been filtered or guided in a detective-led interview resulting in much irrelevant information. We are developing a decision support system (DSS), combining natural language processing (NLP) techniques, similarity measures, and machine learning, i.e., a Naïve Bayes' classifier, to support crime analysis and classify which crime reports discuss the same and different crime. We report on an algorithm essential to the DSS and its evaluations. Two studies with small and big datasets were conducted to compare the system with a human expert's performance. The first study includes 10 sets of crime reports discussing 2 to 5 crimes. The highest algorithm accuracy was found by using binary logistic regression (89%) while Naive Bayes' classifier was only slightly lower (87%). The expert achieved still better performance (96%) when given sufficient time. The second study includes two datasets with 40 and 60 crime reports discussing 16 different types of crimes for each dataset. The results show that our system achieved the highest classification accuracy (94.82%), while the crime analyst's classification accuracy (93.74%) is slightly lower.     

学术社交网络问答质量智能评价与服务优化研究

[目的/意义]学术社交网络所提供的问答服务已成为学者们快速获取学术信息、解决学术问题的重要途径,实现基于机器学习的问答质量智能评价和服务优化对学术社交网络中优质内容传播具有重要意义。[方法/过程]以ResearchGate问答服务为研究对象,从结构化特征、内容特征、其他特征以及回答者特征4个维度构建答案质量评价体系,利用机器学习方法和数据增强技术进行答案质量分类预测。[结果/结论]SMOTE算法在处理不平衡样本时具备有效性;支持向量机在单一模型预测中,取得出色的分类效果;组合模型使预测精度得到进一步提升,基于随机森林、支持向量机、BP神经网络构建的组合模型分类性能最佳,以此为基础可通过搭建问答质量智能评价系统实现学术社交网络问答服务优化。     

2006-2007年国外信息检索基本理论研究进展

2006-2007年国外对信息检索基础理论的研究主要集中于决策理论、隐含语义索引理论研究以及信息检索评价理论研究。关于信息检索基本原理的研究主要集中在信息检索中的分类、信息检索模型、信息检索类型和检索方式等方面。信息检索中的分类的研究重点包括有关分类器的研究;有关特征选择的研究;有关领域相关词的研究。信息检索类型的研究主要包括焦点检索、图像检索、视频检索、合作过滤、机器音译、无线网中网。检索方式的研究主要包括上下文检索、集成检索、问答系统检索以及用户查询处理等问题。     

基于关联数据的敦煌遗书图像知识组织模式的构建

敦煌遗书图像蕴含丰富的文化内涵,对于研究中国古代社会历史、宗教与美术具有重要意义,但传统单一线性的图像检索方式不利于敦煌遗书图像隐性知识的挖掘,影响知识发现的深度与广度。而关联数据能够连接多源异构资源,实现多种资源的语义互联,既能促进管理标准化与规范化,又有利于提升图像内容的深入整合,同时,将关联数据应用于敦煌遗书图像在理论、实践与技术上都具备可行性。为此,本文针对敦煌遗书图像的物理特征与内容语义特征构建敦煌遗书图像层次模型,使用元数据描述后将这些元数据进行关联;同时,本文基于关联数据设计敦煌遗书图像知识关联的组织模式,其自底向上分为数据收集层、语义描述层、数据关联层与知识应用层四层,旨在改善图像检索效果并利于敦煌遗书图像的知识发现与智能查询。     

基于机器学习的人工智能技术专利数据集构建新策略

颠覆性技术是一个具有复杂的内在结构的技术群。从空间维度来看,颠覆性技术是包含了主导技术、辅助技术、支撑技术的复杂技术群,涉及多学科、多领域。在此背景下,运用科学计量的方法对颠覆性技术进行科技评价和科学技术演变规律探索面临挑战,实质表现为数据检索。本文探索了一种基于机器学习的专利数据集构建新策略,将专利检索任务作为机器学习的二分类任务,类似于信息检索中基于主动学习的查询分类思想,并提出了将F-measure特征最大化方法与CNN(convolutional neural networks)模型相结合的文本分类改进方法。本文以人工智能(artificial intelligence,AI)技术域为例进行训练实验,实验结果的准确率、召回率和F1值分别达到98.01%、97.04%和97.89%,这表明本文提出的策略能够精准地识别人工智能专利,提高了专利检索的准确率和召回率,以利于构建精、准、全的人工智能技术域专利数据集。     

Investigation of the Painting Idyll Attributed to Henryk Siemiradzki: The Unusual Technology of a Canvas Painting Executed on an Enlarged Photograph

The subject of the research is an oil painting entitled Idyll created using an unusual technique. The painting layer was applied on canvas covered with photosensitive medium on which a photograph was developed. The results of the examination of Idyll in the context of its attribution to Henryk Siemiradzki (1843–1902) are discussed. Its atypical technology and the possibility of the artist having used the photographic technique are also considered. X-ray fluorescence spectroscopy, infrared spectroscopy, and scanning electron microscopy were performed to establish the technique and to characterize the pigments and materials used in the painting. The identified pigments and the information provided by analytical photography (UV, X-ray images) were compared with the results of the detailed examination of 56 of Siemiradzki’s oil paintings. Performed analyses revealed that the majority of the features identified in Idyll point to Henryk Siemiradski’s technique.     

西方画作中的阅读史

[目的/意义] 以西方画作为切入点研究西方阅读史的发展历程,探究图像视角下西方阅读史的发展特点,实践以图证史的研究路径,为中国阅读史的研究提供借鉴。[方法/过程] 以588幅阅读主题的西方画为研究对象,使用图像分析法,在梳理和描绘图像视角下西方阅读史发展演变的基础上,分析不同时代对阅读行为认知的演变,最后总结图像视角下的西方阅读史特征。[结果/结论] 图像视角下的西方阅读史具有以人物为中心进行叙述的特征,宗教主题的阅读贯穿始终。使用绘画作品进行阅读史研究应该考虑时代因素、绘画目的和艺术加工程度的影响。     

The effects of atmospheric moisture on the mural paintings of the Mogao Grottoes

The aim of this paper is to study the influence of atmospheric humidity and temperature on the mural paintings in the Mogao Grottoes of Dunhuang by measuring the weight of a simulated mural block. Under open conditions, the daily changes in the atmospheric humidity and temperature have an apparent effect on the water content of these murals. There exists an obvious water exchange between atmosphere and mural, that is, there is a ±43?g?m^?2 moisture absorption–desorption ‘breathing’ process between the two. Evapotranspiration from the tree-belt, precipitation, and extremely dry weather also have an effect on the moisture associated with a mural painting. If the cave is closed, a comparable study finds that closure can make the effects of temperature and humidity changes disappear. The mural water content in this case remains stable and the harm due to water-salt deterioration is greatly reduced. Under closed conditions, artificial condensation dehumidification and control of the cave's temperature and humidity stabilizes water activity in the mural paintings very effectively. This is a clear indication of the future steps required to protect the cave's cultural relics.