文本风格迁移研究综述

文本风格迁移是近年来自然语言处理领域的热点问题之一,旨在保留文本内容的基础上通过编辑或生成的方式更改文本的特定风格或属性(如情感、时态和性别等).旨在梳理已有的技术,以推进该方向的研究.首先,给出文本风格迁移问题的定义及其面临的挑战;然后,对已有方法进行分类综述,重点介绍基于无监督学习的文本风格迁移方法并将其进一步分为隐式和显式两类方法,对各类方法在实现机制、优势、局限性和性能等方面进行分析和比较;同时,还通过实验比较了几种代表性方法在风格迁移准确率、文本内容保留和困惑度等自动化评价指标上的性能;最后,对文本风格迁移研究进行总结和展望.     

Layout extraction of mixed mode documents

Proper processing and efficient representation of the digitized images of printed documents require the separation of the various information types: text, graphics, and image elements. For most applications it is sufficient to separate text and nontext, because text contains the most information. This paper describes the implementation and performance of a robust algorithm for text extraction and segmentation that is completely independent of text orientation and can deal with text in various font styles and sizes. Text objects can be nested in nontext areas, and inverse printing can also be analyzed. It should be mentioned that the classification is based only on rough image features, and individual characters are not recognized. The three main processing steps of the system are the generation of connected components, neighborhood analysis, and generation of text lines and blocks. As output, connected components are classified as text or nontext. Text components are grouped as characters, words, lines, and blocks. Nontext objects are accumulated as a separate nontext block.     

A robust arbitrary text detection system for natural scene images

Text detection in the real world images captured in unconstrained environment is an important yet challenging computer vision problem due to a great variety of appearances, cluttered background, and character orientations. In this paper, we present a robust system based on the concepts of Mutual Direction Symmetry (MDS), Mutual Magnitude Symmetry (MMS) and Gradient Vector Symmetry (GVS) properties to identify text pixel candidates regardless of any orientations including curves (e.g. circles, arc shaped) from natural scene images. The method works based on the fact that the text patterns in both Sobel and Canny edge maps of the input images exhibit a similar behavior. For each text pixel candidate, the method proposes to explore SIFT features to refine the text pixel candidates, which results in text representatives. Next an ellipse growing process is introduced based on a nearest neighbor criterion to extract the text components. The text is verified and restored based on text direction and spatial study of pixel distribution of components to filter out non-text components. The proposed method is evaluated on three benchmark datasets, namely, ICDAR2005 and ICDAR2011 for horizontal text evaluation, MSRA-TD500 for non-horizontal straight text evaluation and on our own dataset (CUTE80) that consists of 80 images for curved text evaluation to show its effectiveness and superiority over existing methods.     

iPad-assisted reading fluency instruction for fourth graders with reading difficulties: A single case experimental design

The purpose of this study was to investigate the effect of iPad-assisted reading instruction for fourth-grade students' oral reading fluency and their attitudes towards reading. Three fourth-grade students at-risk for reading disabilities participated in an iPad-assisted repeated reading programme that integrated repeated reading, vocabulary instruction, error correction and feedback procedures. During each intervention session, the participants practiced reading a fourth grade-level text, engaged in repeated reading with the intervention agent and repeatedly read the text three or four times with error correction. The oral reading rate for the fourth-grade text that they practiced and a transfer text (i.e., new text that students have not practiced before) served as the primary dependent variables. Their attitudes towards reading were measured before and after the intervention. Using a multiple baseline across participants design, we showed that the iPad-assisted repeated reading programme improved all participants' oral reading rates on both the grade-level text and the transfer text. Implications of the study and future directions were discussed.     

基于汉字连通分量的印刷图像版面分割方法

针对彩色印刷图像背景色彩丰富和汉字存在多个连通分量,连通域文字分割算法不能精确提取文字,提出基于汉字连通分量的彩色印刷图像版面分割方法。利用金字塔变换逆半调算法对图像进行预处理,通过颜色采样和均值偏移分割图像颜色,标记文字连通分量,根据汉字结构和连通分量特性重建汉字连通分量,分析文字连通分量连接关系确定文字排列方向实现文字分割。实验结果表明,该方法能够有效地重建汉字连通分量,在彩色印刷图像中实现对不同字体、字号、颜色的文字分割。     

结合迁移引导和双向循环结构GAN的零样本文本识别

为了提高基于生成对抗网络(GAN)的零样本识别方法的识别精度,提出结合迁移引导和双向循环结构GAN的零样本文本识别方法.构造双向循环结构GAN以提高模型的生成能力,生成的伪特征更接近输入的真实特征.引入迁移引导学习的思想,使用迁移后的文本代替可见类文本训练模型,提高不可见类文本的识别精度.增加有效的正则化项,使生成器在训练过程中生成的结果具有多样性,提高生成模型的稳定性.在数据集上的实验表明,文中方法可提高识别精度,具有较好的泛化性能,容易拓广到其它应用中.     

融合迁移学习和神经网络的皮肤病诊断方法

针对医学特征对患者病情发展的时间顺序无法有效表达,医学特征构建工作耗费大量人工成本,以及皮肤病数据样本数量较少等问题,提出了融合迁移学习和神经网络的皮肤病辅助诊断方法。该方法将TextLSTM(long short term memory neural network for text)、TextCNN(convolutional neural network for text)以及RCNN(recurrent convolutional neural networks for text classification)等3种基于神经网络的文本分类模型应用于皮肤病辅助诊断,同时融入迁移学习技术,能够在一定程度上将皮肤病专业书籍中的理论知识迁移到诊断模型中。在皮肤病多分类实验中,本文方法的正确率优于对比方法;在皮肤病二分类实验中,本文方法的召回率优于对比方法。迁移学习对实验结果的积极影响率高于75%。     

基于有限元的PCB板上关键元件热可靠性分析

电子设备不断地微型化,热设计就显得越来越重要。体积小、布局紧凑,导致元件温升越高,从而大大降低系统的可靠性。为此文章从热传输原理出发,运用ANSYS有限元软件分析印刷电路板(PCB)上关键元件工作时的温度场分布,确定PCB的高温区和低温区。并通过实例计算不同布局的PCB的温度场,通过比较得出较为合理布局方式。优化布局,降低PCB板的最高温度,提高系统的可靠性。     

基于ResNet和迁移学习的古印章文本识别

古印章文本因图像退化与超多分类等特点导致识别难度大,部分字符的标注数据不足造成基于深度学习的模型识别准确率不高,泛化能力差.针对上述问题,提出基于深度残差网络(ResNet)和迁移学习的古印章文本识别方法.使用深度残差网络作为特征提取网络,利用人工合成字符样本作为源域进行预训练.将自建古印章文本识别数据集作为目标域,引...     

车牌图像色彩迁移与正则化约束去雾算法

针对雾天车牌图像模糊、车牌识别率低的问题,给出了车牌图像色彩迁移与正则化约束去雾算法。算法主要包含色彩迁移去雾和文本修复两个模块。采用MKL（Monge-Kantorovitch Linear Colour Mapping）色彩迁移算法,恢复雾天车牌颜色信息实现去雾;利用车牌的文本像素的强度和梯度特征对车牌图像进行正则化约束,实现车牌中文本的修复。实验结果表明,无论针对合成车牌雾图还是自然车牌雾图,去雾效果良好,且在薄雾、中等雾及浓雾三种不同雾度环境下都能够有效提高车牌识别率。     

EFBS数据交换模型与完整性检查机制

提出一种面向电子飞行包系统(EFBS)的数据交换模型。该模型采用星型逻辑结构进行数据传输,以XML格式通过数据交换组件实现其他子系统与中央内容管理单元的数据交换,介绍数据完整性检查机制,确保数据交换的安全性,其中,发送方通过DSA算法生成数字签名,并运用SHA-1算法获得文本摘要,接收方通过数字签名验证对交换数据进行完整性检查。实验结果验证了该检查机制的有 效性。     

基于文本重要内容的鲁棒水印算法

提出一种基于文本重要内容的鲁棒水印算法, 对文本的特征进行分析, 确定文本的重要内容. 根据水印序列和同义词替换评价模型, 将水印不可感知地嵌入到文本的重要内容中, 提高水印的鲁棒性. 投票原则的使用又进一步提高水印的鲁棒性并降低了误检率. 理论和实验分析表明, 采用文中算法嵌入的水印具有较好的鲁棒性、安全性和不可见性.     

脱机手写维吾尔文本图像单词切分

针对脱机手写维吾尔文本行图像中单词切分问题,提出了FCM融合K-means的聚类算法。通过该算法得到单词内距离和单词间距离两种分类。以聚类结果为依据,对文字区域进行合并,得到切分点,再对切分点内的文字进行连通域标注,进行着色处理。以50幅不同的人书写的维吾尔脱机手写文本图像为实验对象,共有536行和4?002个单词,正确切分率达到80.68%。实验结果表明,该方法解决了手写维吾尔文在切分过程中,单词间距离不规律带来的切分困难的问题和一些单词间重叠的问题。同时实现了大篇幅手写文本图像的整体处理。     

一种视频中字符的集成型切分与识别算法

视频文本行图像识别的技术难点主要来源于两个方面: 1)粘连字符的切分与识别问题; 2)复杂背景中字符的切分与识别问题. 为了能够同时切分和识别这两种情况中的字符, 提出了一种集成型的字符切分与识别算法. 该集成型算法首先对文本行图像二值化, 基于二值化的文本行图像的水平投影估计文本行高度. 其次根据字符笔划粘连的程度, 基于图像分析或字符识别对二值图像中的宽连通域进行切分. 然后基于字符识别组合连通域得到候选识别结果, 最后根据候选识别结果构造词图, 基于语言模型从词图中选出字符识别结果. 实验表明该集成型算法大大降低了粘连字符及复杂背景中字符的识别错误率.     

基于注意力机制的卷积-双向长短期记忆模型跨领域情感分类方法

针对现有跨领域情感分类方法中文本表示特征忽略了重要单词的情感信息，且在迁移过程中存在负迁移的问题，提出一种基于注意力机制的卷积-双向长短期记忆（AC-BiLSTM）模型的知识迁移方法。首先，利用低维稠密的词向量对文本进行向量表示；其次，采用卷积操作获取局部上下文特征之后，通过双向长短期记忆（BiLSTM）网络充分考虑特征之间的长期依赖关系；然后，通过引入注意力机制考虑不同词汇对文本的贡献程度，同时为了避免迁移过程中出现负迁移现象，在目标函数中引入正则项约束；最后，将在源领域产品评论训练得到的模型参数迁移到目标领域产品评论中，并在少量目标领域有标注数据上进行微调。实验结果表明，与AE-SCL-SR方法和对抗记忆网络（AMN）方法相比，AC-BiLSTM方法的平均准确率分别提高了6.5%和2.2%，AC-BiLSTM方法可以有效地提高跨领域情感分类性能。     

The intermediary language for multilanguage translation

This paper discusses some basic notions involved in designing, developing, and implementing the Intermediary Language (IL) for Machine Translation applied to a set of languages. The stages for the design of the IL would include the independent analysis and synthesis of each language in its own terms. Then each could be mapped once into the IL dictionary and grammar, creating the IL text. From the IL text the transfer routine would synthesize the target text for a particular language. It is assumed that the IL text would have algebraic representation of the variables to be instantiated in the target language on the basis of the IL text information. The IL should contain all the information occurring in the set of languages plus such generalizations as might be justified on the basis of inductive implications and/or deductively oriented postulates to be verified by adding new languages for testing the capacity of the IL.Given five languages spoken by more than a hundred million people, if N equals 5 for the pairwise translation (say, into English), N²–N, we get 20 programs, and for the IL translation 2N+1, we can manage with eleven programs, yielding a significant gain.The IL metalanguage, ideally, should have the capacity to function as an algebraic representation of both paradigmatic units (the selection axis) and their relationships (the contiguity axis). Both should be correlated with the extralinguistic fragments in terms of determiners, quantifiers, and classifiers. The structure of the IL grammar contains four components: dictionary, context-free information providing the nonterminal dictionary (i.e., classification), parser/synthesizer, and the initial string.     

基于小波形态学的文本自动检测

视频数据中的文本能提供重要的语义信息。论文针对视频中的文本分割问题,提出一种分割方法,用具有良好时频局部和变尺度特性的小波分析方法,提取出纹理清晰、具有不同空间分辨率、不同方向的边缘子图象,然后,用数学形态学的方法对提取出的细节图象进行进一步的形态运算,消除无用的信息,得到最终的文本目标区域。实验表明该方法简单有效,适合于背景较复杂的文本分割。     

Accurate video text detection through classification of low and high contrast images

Detection of both scene text and graphic text in video images is gaining popularity in the area of information retrieval for efficient indexing and understanding the video. In this paper, we explore a new idea of classifying low contrast and high contrast video images in order to detect accurate boundary of the text lines in video images. In this work, high contrast refers to sharpness while low contrast refers to dim intensity values in the video images. The method introduces heuristic rules based on combination of filters and edge analysis for the classification purpose. The heuristic rules are derived based on the fact that the number of Sobel edge components is more than the number of Canny edge components in the case of high contrast video images, and vice versa for low contrast video images. In order to demonstrate the use of this classification on video text detection, we implement a method based on Sobel edges and texture features for detecting text in video images. Experiments are conducted using video images containing both graphic text and scene text with different fonts, sizes, languages, backgrounds. The results show that the proposed method outperforms existing methods in terms of detection rate, false alarm rate, misdetection rate and inaccurate boundary rate.     

Using semantic components to search for domain-specific documents: An evaluation from the system perspective and the user perspective

We seek to leverage an expert user's knowledge about how information is organized in a domain and how information is presented in typical documents within a particular domain-specific collection, to effectively and efficiently meet the expert's targeted information needs. We have developed the semantic components model to describe important semantic content within documents. The semantic components model for a given collection (based on a general understanding of the type of information needs expected) consists of a set of document classes, where each class has an associated set of semantic components. Each semantic component instance consists of segments of text about a particular aspect of the main topic of the document and may not correspond to structural elements in the document. The semantic components model represents document content in a manner that is complementary to full text and keyword indexing. This paper describes how the semantic components model can be used to improve an information retrieval system. We present experimental evidence from a large interactive searching study that compared the use of semantic components in a system with full text and keyword indexing, where we extended the query language to allow users to search using semantic components, to a base system that did not have semantic components. We evaluate the systems from a system perspective, where semantic components were shown to improve document ranking for precision-oriented searches, and from a user perspective. We also evaluate the systems from a session-based perspective, evaluating not only the results of individual queries but also the results of multiple queries during a single interactive query session.     

A reversible information hiding scheme using left-right and up-down chinese character representation

Techniques for text data hiding are different from image data hiding, video data hiding and audio data hiding. To break through the difficulty of text data hiding, Sun, Lou and Huang proposed a novel Chinese text data hiding scheme called the L-R scheme. In the L-R scheme, Sun et al. embedded secrets into Chinese characters that can be divided into left and right components. This paper describes how our proposed scheme extends the component concept to incorporate the up and down components of Chinese characters rather than the left and right components only, to significantly enhance hiding capacity. In addition, this paper adds a reversible function to Sun et al.’s L-R scheme to make it possible for receivers to obtain the original cover text and use it repeatedly for later transmission of secrets after the initial hidden secrets have been extracted. Finally, the extended scheme simplifies the extracting procedure and efficiently reduces the memory required on the receiver side during the secret extracting phase by using a new comparison method. Experimental results confirm the improved functions offered by the proposed scheme.